Beyond Imitation Learning: How E2E Self-Play Reinforcement Learning Scales for Autonomous Driving

Beyond Imitation Learning: How E2E Self-Play Reinforcement Learning Scales for Autonomous Driving

Beyond Imitation Learning: How E2E Self-Play Reinforcement Learning Scales for Autonomous Driving (Draft)

How self-play DAgger distills a privileged Reinforcement Learning driving policy into a camera-based policy

Presenting at CoRL 2026 in Austin, TX, November 9–12.

TL;DR

  • On the highway, the situations that matter most are the ones least likely to be represented in human driving logs. Models that learn by imitating those logs eventually stop improving.
  • Self-play reinforcement learning lets a driving policy generate its own experience. Our paper, Scaling Self-Play for End-to-End Driving, written with Mila, NYU and collaborators and to be presented at CoRL 2026, shows how to do that for end-to-end models that drive from camera images.
  • Robust driving doesn’t need photorealism: our simplified simulator renders about 1,000× faster than a photorealistic one. The resulting camera-based policy, trained without human driving trajectories, keeps improving where imitation plateaus and scores best among the methods evaluated on the closed-loop benchmark.

The data runs out where it matters most

At 100 km/h, a truck covers close to 30 meters every second. Most of those seconds are uneventful. The ones that decide whether a driver is good (a stalled car in the lane ahead, a cut-in at speed, a merge that closes faster than expected) are rare on the road, so they are rare in recorded data too.

Most end-to-end driving models today are trained with imitation learning, also called behavior cloning: the model is shown recorded human driving and learns to copy what the human did in each situation. It never drives on its own during training and never gets feedback on the outcome of its own choices. That is what separates it from reinforcement learning.

A model trained to imitate human driving learns the common case very well and the rare case barely at all. It also never practices recovering from its own mistakes, because a log only contains what the human did. Once the model drives on its own, a small error puts it somewhere the data never went, and the next error is bigger.

Recording more miles helps less than you might expect, since most new miles are more examples of driving that was already going fine. In our experiments, an end-to-end model trained by imitation stopped improving at around 100 million agent steps, where one agent step is one vehicle taking one action in the simulator.

The same end-to-end model trained three ways, tested in closed loop. Imitation flattens out; self-play keeps climbing.

Experience you generate, not data you collect

Reinforcement learning (RL) takes a different route. Imitation learning copies what a human did; RL lets the model drive in a simulator, try actions, and learn from a reward that scores what happened: progress made, collisions avoided, a smooth ride. Self-play reinforcement learning is a specific kind of RL in which the model doesn’t drive against recorded traffic. Instead, one policy drives every vehicle in a simulated scene and learns from all of them at once. Because everyone on the road is learning, the traffic keeps getting harder in the ways that matter: gaps close faster, merges get tighter, the car ahead brakes. The rare situations stop being rare, because the agents keep creating them for each other.

Training one vehicle against recorded traffic, versus self-play, where one policy drives everyone.

With self-play, the ceiling moves from how much data you can collect to how much simulation you can run.

So why isn’t self-play reinforcement learning already the standard way to train driving models? So far, it has worked mainly for policies that see the world as a structured list of objects and lanes. End-to-end models that drive from camera images face two obstacles. First, every self-play step needs fresh camera images, and photorealistic simulators produce them roughly a thousand times slower than the renderer we describe below. Second, RL is sample-inefficient: its policy updates need billions of agent steps of experience, and with a large vision model, every step costs a forward and a backward pass. This work addresses both.

Three ideas that make it work

You don’t need photorealism to learn to drive

What a driver needs from a scene is where things are and how they move, not what they look like. So we built Gigapixel, a simulator that draws the world as plain boxes for vehicles, thin strips for lanes and small spheres for traffic lights, and renders thousands of these worlds at once on a GPU, fast enough for self-play reinforcement learning from pixels. Its renderer sustains about 50,000 agent steps per second on a single NVIDIA A100 GPU, roughly 1,000 times faster than a photorealistic reconstruction simulator. To put that in perspective: the camera-based student in this work trained on 150 million agent steps. At Gigapixel’s rate, one GPU renders all of them in under an hour; at the speed of the photorealistic simulator we compared against, the same rendering would take more than a month. Throughput also scales nearly linearly as GPUs are added.

What the policy sees during self-play: the forward camera views stitched together. Blue boxes are other vehicles, thin strips are lane lines, and small spheres are traffic lights. Videos from the Gigapixel project page.

Learn with RL where it’s cheap; distill where it’s expensive

Reinforcement learning needs billions of agent steps of simulated experience; the teacher in this work trained for 25 billion. That is affordable for a small policy that reads a list of objects, and far too costly for a large vision model that has to process four camera images at every step. So we split the job into two steps. Step one is RL: it trains a compact teacher that sees the scene as structured data, a list of objects and lanes, which is cheap to simulate. Step two is the distillation: instead of training the camera-based student with RL, we train it with supervised learning to match the teacher’s planned trajectories. The student drives every vehicle in the simulation, and wherever it ends up, the teacher shows it what it should do from there. This technique is known as DAgger; because we run it in self-play, we call it self-play DAgger (see panel (b) in the figure below). Because the student learns in the situations it actually creates, including its own mistakes, it learns to recover from them. It reaches the same level of driving in roughly 3,000 times fewer agent steps than training the camera model with RL directly.

Only the eyes need to adapt

A student trained on boxes still has to see real roads. The skill of driving lives in the part of the network that plans, so we leave that part alone and retrain only the part that turns images into an internal picture of the scene. Using pairs of real camera frames and renderings of the same moment, we teach it to see a real road the way it saw the simulated one. No human driving trajectories are used at any stage.

How the pieces fit together

The figure below, from the paper, shows the full training pipeline, from left to right.

The training pipeline: (a) Vectorized Self-Play RL, an RL teacher on structured inputs; (b) Pixel-based Self-Play DAgger in Gigapixel, a camera-based student distilled from the teacher in self-play; and (c) Sim-to-Real Perception Adaptation to real camera images. Figure 2 from the paper.
  • The teacher: Vectorized Self-Play RL. A compact network of about 2.7 million parameters, trained with reinforcement learning (PPO algorithm) in self-play for 25 billion agent steps across 335,000 driving scenarios from nuPlan. Its reward weights for things like collisions, comfort and progress are randomized for each vehicle and given to the policy as an input. One model therefore learns a whole range of driving styles, from cautious to assertive, which gives the student varied traffic to drive against.
  • The student: Pixel-based Self-Play DAgger in Gigapixel. A transformer-based E2E planner (adapted from the DrivoR paper) that sees four rendered camera views and outputs a 4-second trajectory. It controls every vehicle in the scene. At each situation it reaches, the simulator is copied and the teacher drives on from that exact point for 4 seconds; that rollout becomes the student’s target. Early in training the teacher still takes some of the actions, and control passes fully to the student over time, which keeps self-play stable while the student is still learning.
  • The adaptation: Sim-to-Real Perception Adaptation. The trained student is duplicated. One copy looks at the rendering of a scene, the other at the matching real camera frame, and only the real-image copy’s perception backbone is updated, so that it produces the same internal features and the same planned trajectory. The planning head is frozen throughout. Both choices matter: without the feature-matching term, the HUGSIM score drops from 33.2 to 18.5.

The full method, ablations and benchmark tables are on the project page and in the paper.

What we learned

Self-play changes what data you get, not just how much

This is where the approach answers the problem we started with. Human driving logs are thin exactly where it matters: the stalled car, the cut-in at speed, the merge that closes too fast. The obvious benefit of self-play reinforcement learning is volume: every vehicle in a rollout becomes a training example, so each simulated second yields many. The bigger benefit is content. Because every driver is learning, the scenarios shift as training goes on, and the student keeps running into interactions that a fixed dataset would show it once, if ever. Self-play turns the sparse long tail of the logs into a dense, steady supply of exactly the situations that improve driving behavior. That is why it beat single-agent training at every scale we tested and kept improving past the point where imitation stalled.

You can see the difference in behavior. When a lead vehicle slows to a stop, the self-play policy eases off and plans a smooth path around it; the imitation-trained policy keeps going straight and rear-ends it.

Same scene, same model architecture, different training. More examples on the project page.

“Imitation learning is limited by the sparsity of useful long-tail driving scenarios, so it tends to plateau in performance. Self-play reinforcement learning greatly increases the density of these high-value scenarios to avoid this plateau, showing that it can enable much more robust, safe and performant planning that scales.”

Felix Heide
Head of AI, Torc Robotics
Professor at Princeton


A higher score isn’t the same as safer driving

We evaluated our approach on the HUGSIM benchmark, which reconstructs real driving scenes in a photorealistic simulator and lets a policy drive through them in closed loop, with surrounding traffic that becomes more adversarial across four tiers, from Easy to Extreme.

Overall, our self-play policy scored best among the methods evaluated in the paper, without using any human trajectories.* The exception was the Extreme tier, where surrounding vehicles actively try to cause collisions. There, the imitation-trained version of the same model scored higher.

Inspecting the runs showed why. The imitation-trained model tends to drive fast, which lets it outpace the aggressive traffic before a collision happens. Ours drove more cautiously, yielded to those vehicles and often got stuck, which lowers route completion and, with it, the score. Across the full evaluation set, the imitation-trained model’s average speed at collision was 2.7 times higher than ours.

Average speed at collision on HUGSIM (lower is better)

Same DrivoR architecture, trained two ways; averaged over the full HUGSIM evaluation set. Source: paper, Section 4.
0246DrivoR planner, same architectureImitation (behavior cloning)5.27 m/sSelf-play DAgger1.95 m/s
Trained by imitating human drivingTrained with self-play
Average speed at the moment of collision for the same DrivoR-based planner trained two ways, across the full HUGSIM evaluation set. When collisions occurred, the imitation-trained model was moving at 5.27 m/s on average versus 1.95 m/s for the self-play-trained model, a 2.7× reduction. Data from the paper, Section 4.

For a heavy truck, that trade-off isn’t close. At the same speed, a loaded truck carries far more energy into a collision than a passenger car, so a policy that sometimes waits behind aggressive traffic shows much safer behavior.

What benchmarks can and can’t tell you

Benchmarks like HUGSIM are valuable in a research setting. They let us compare methods on equal terms and check, at a high level, whether an idea holds up outside our own simulator. But as the example above shows, a single aggregate number can’t fully capture good driving behavior, and it can even reward the wrong behavior. That is why at Torc we evaluate how a policy drives from many perspectives and at different levels of depth, using a wide range of metrics and observers, both in simulation and in real-world truck tests.

The renderer stopped being the bottleneck

Once a full-size end-to-end model is in the training loop, most of the cost is the model itself, not drawing the images. That leads to the most practical lesson of this work: the best simulator for training isn’t necessarily the most realistic one; it is the one that keeps the model learning from as many useful situations as possible per GPU-hour. That is what self-play reinforcement learning in a simple box world gives us. It generates a much higher density of rare, useful scenes than any log can, and that keeps the model learning well past the point where imitation plateaus. What is learned in the box world doesn’t stay there, either: the self-play RL policy trained on structured inputs distills efficiently into an end-to-end policy that drives from camera pixels, and a light perception adaptation carries that behavior over to real camera images.

What a box world can’t teach

Boxes and lane strips can’t represent debris, unusual obstacles, weather or lighting, so the policy never practices those in simulation. The student also has to infer from pixels what the teacher was simply given, which isn’t always possible when something is hidden. And the adaptation step needs paired real and rendered views. These limitations also show where the research goes from here: adding human driving data back into training so policies stay human-like, generating rarer and more informative starting scenarios, and eventually training camera-based policies with RL directly, without a teacher.

Why this matters for Torc Robotics

Highway trucking raises the stakes on everything above. Popular urban datasets concentrate on objects within 100 m and speeds under 50 km/h; long-range highway data such as the TruckDrive dataset extends labels to 400 m, speeds above 100 km/h, and trajectories to 800 m. At that range and speed, the interactions that matter are the rarest in any log, which is exactly where generated experience should pay off most. Trucks also move differently from cars: this work was evaluated on passenger-car benchmarks, and adapting it to heavy, articulated vehicles is part of the work ahead.

This research also connects directly to what we run today. The teacher in this paper is the same kind of structured self-play reinforcement learning policy we use in our production work, and end-to-end RL is where our research is heading. The paper shows one path between the two: keep the robustness that self-play builds, and carry it into models that drive directly from cameras. We’ll share more about RL on our trucks in an upcoming post.

Ongoing collaboration with Mila, NYU and partners

None of this happens alone. The work is a collaboration between Torc, Mila, NYU’s EMERGE Lab and partner universities, and it builds on open-source tools: Gigapixel extends PufferDrive, and its code is coming soon. If these are the problems you want to work on, we’re hiring.

Meet the Torc Robotics RL team at CoRL 2026

Poster Session 3 | Poster 207
Wednesday, November 11, 2026 • 10:30 AM–12:00 PM • JW Marriott

We’ll be presenting Scaling Self-Play for End-to-End Driving as a poster at the Conference on Robot Learning, and members of our RL team will be there to meet. Come find us to talk about:

  • Self-play reinforcement learning and closed-loop training for driving, from structured policies to end-to-end models
  • How we take RL from research into engineering for autonomous trucks
  • Open roles on our RL and ML behaviors, End-2-End, Simulation and ML infrastructure teams

Go deeper

Paper: Scaling Self-Play for End-to-End Driving (CoRL 2026, Austin, TX)
Project page, with videos, full results and method details
Code (coming soon)

Authors

Luke Rowe (Mila, Université de Montréal), Roger Girgis (Mila, Polytechnique Montréal, Torc Robotics), Rodrigue de Schaetzen (Mila, Université de Montréal, Torc Robotics), Daphne Cornelisse (NYU Tandon), Alaap Grandhi (Mila, McMaster University), Felix Heide (Torc Robotics, Princeton University), Eugene Vinitsky (NYU Tandon), Christopher Pal (Mila, Université de Montréal, Polytechnique Montréal), Liam Paull (Mila, Université de Montréal)

Related Research

What is DL+S? Unlocking the Massive Data Required to Develop Autonomous Trucks

What is DL+S? Unlocking the Massive Data Required to Develop Autonomous Trucks

Graphic featuring Torc truck with code overlay

Breakthroughs in AI for autonomous driving do not happen in isolation. They happen when data, simulation, machine learning, infrastructure, and vehicle testing work as one integrated engine for innovation. Recognizing that, Torc created the DataLoop and Simulation (DL+S) division in the AI organization to connect every link in the development cycle and accelerate progress at scale.

As Torc progressed toward its AV 3.0 vision, we recognized the need to unify data ingestion, data processing, experimentation, training, validation, and deployment across the organization. Achieving that scale required both organizational alignment and technical infrastructure capable of orchestrating massive machine learning workloads. The maturing of an autonomous driving organization required an investment in centralization and in standardized tooling and led Torc to standardize on technologies such as Ray by Anyscale as part of its broader DataLoop strategy.

This approach has required substantial upfront investment and cultural buy-in, but the long-term benefits are considerable. We have already seen how standardized workflows improve development velocity by enabling faster support and maintenance code, thereby reducing duplication of effort. Common tooling enables better auditability and repeatability, which are essential for safety-critical systems. Shared platforms also allow teams to collaborate more effectively instead of rebuilding similar capabilities multiple times.

The organization supports the complete lifecycle of autonomous driving development through what Torc refers to as the DataLoop.

Understanding the DataLoop

At its core, the DataLoop is a continuous cycle of learning, validation, and improvement.

Closed-Loop Autonomous Truck Data Lifecycle

The process begins with data ingest, where new information arrives from vehicle operations, testing, and other sources. Each Torc truck mission produces roughly 9TB/hour of new data, all of which must be managed in perpetuity to satisfy safety standards. That raw data then moves through data processing, where it is supplemented with the appropriate for use in machine learning and simulation.

Next comes model development and model validation, where engineers train and refine AI models as part of model experimentation. Experimentation is not linear, but instead a loop that constantly uses new insights gained from model validation to guide development.  Autonomy teams continually evaluate results, adjust model architecture and hyperparameters, augment datasets, and repeat the process to improve performance. DL+S is responsible for accelerating this experimentation loop, both through faster computation and better UI/UX to surface and share developer insights.

Once a model demonstrates promising results, it enters convert & deploy, where it is prepared for operation on our production hardware. This requires an auditable workflow from TensorRT and eventually a TorcDrive cwhich can undergo final performance and latency validation. Finally, a new stack release which is deemed performant and safe undergoes truck testing, which renders the final safety validation step. We both celebrate our progress and surface.

How Much is 9TB/Hour of Data?

Trying to figure out how much 9TB of data is?

9 TB is about 2,250,000 songs worth of data, or 2.5 million standard smartphone photos. And it’s almost a tenth of the entire printed collection of the United States Library of Congress.

Torc trucks are producing that amount of data per hour.

“It is our job, at DataLoop + Sim, to recognize these challenges, implement scalable solutions, and partner with our Autonomy colleagues to drive the technology forward.”

Charles Molnar

Head of DL+S, Torc

What Makes Autonomous Trucking Difficult?

Autonomous trucking presents a unique challenge because of the sheer scale and complexity of the data involved, as well as the operational domain in which freight shipments must succeed.  Felix Heide, Torc’s head of AI, recently spoke about this at Ray Summit; he reminded the audience that guiding a vehicle weighing up to 80,000 pounds, packed in an unknown fashion, moving at 65 miles/hour is a very different task than non-trucking autonomous vehicle benchmarks. Simultaneously, unlike conventional machine learning settings, autonomous driving systems must process and learn from massive multimodal datasets. Camera imagery, lidar point clouds, radar measurements, audio signals, vehicle telemetry, and additional inputs all must be combined into a coherent understanding of the world.  Training on this multimodal data requires elaborate data management to efficiently utilize the heterogeneous compute infrastructure that Torc uses.  The compute challenge is not simply acquiring large amounts of compute: that is necessary, but not sufficient. Torc succeeds when DL+S enables the effective use of compute. Our challenge starts before the first scrap of truck data is acquired: we need to have a storage system that can manage 100+ PB of data. Then:
  • We need to design and implement a data format that is both rigid enough to scale, but also flexible enough to accommodate future sensor/hardware changes.
  • We need auto-labeling processes, hand-checked by our manual labeling team, to identify useful data and process it for training.
  • We need training frameworks that can effectively balance the compute task to the right computer power, to train on this data in hours, not days.
  • Lastly, we need simulation, at all levels of fidelity for truck-specific motion models, to both power our reinforcement learning process and confirm that the model we trained will have suitable post-conversion performance as a deployable TorcDrive component.

The Role of Simulation Throughout the Loop

Simulation is embedded throughout several stages of the DataLoop, which is why it is called out separately in the division’s name.

DL+S uses simulation to support synthetic data generation, enabling teams to create validation scenarios that may be difficult, expensive, or time-consuming to capture in the real world. Simulation is the engine of reinforcement learning, allowing models to learn from large volumes of virtual interactions and gain hundreds of years of driving experience in hours. And, for all autonomy teams, model validation through replay and recompute workflows helps developers analyze prior vehicle behavior, compare against changes, and test improvements under controlled conditions.

Especially in the context of model validation, every investment in improving our simulation capabilities gives us more ability to operate efficiently.  Higher fidelity models give better estimates of truck performance and maximize the value of physical truck launches and tests by excluding new model versions which are insufficiently performant.  Lower fidelity models offer faster ability to perform simulation and explore the domain of possible circumstances worth further study.

Compilation of rendered scenes training our models,

Four Teams

Torc’s DL+S organization brings together four tightly connected teams:

Simulation

Simulation is, unsurprisingly, responsible for building our in-house sim tooling and ensuring its successful use in synthetic data generation, reinforcement learning, and validation.  This requires both research into high fidelity truck motion simulation and implementation of differentiable simulation to facilitate training at a pace approaching 1 million steps/second.

Data Ops

Data Ops is responsible for all of the ingestion and maintenance of truck data at Torc.  When we drive a data collect mission, buy a dataset, or generate synthetic data, or when we add labels to data, either manually or with auto-labeling models, they manage it.  When we conduct a simulation, they store the outputs and make them available for subsequent processes.

ML Ops

ML Ops has the broad mandate of working to ensure compute infrastructure exists for the entire AI Org, alongside the appropriate frameworks for efficient usage of that infrastructure.  This includes training computer vision models, reinforcement learning, batch inference, KPI dashboarding, and model conversion.

ML Data
ML Data builds and maintains tooling for studying and operating on data using ML or data science.  They build the auto-labeling models which take raw data and enable training on it.  They also build VLM-based models for scenario identification on raw data, such as the presence of rain, or police officers, or traffic cones.

The DataLoop for Today and the Future

By standardizing tooling, improving auditability, enabling repeatable experimentation, and supporting ever-larger multimodal AI workloads, DL+S transforms individual engineering efforts into a unified development system. As Torc continues advancing toward fully autonomous trucking, the DataLoop serves as the engine that continuously converts new data into better performance on the road. We strive to maximize learning/minute.

But the success of DL+S demands more than digital infrastructure. Accelerating development is not a purely technological goal, it is also a culture goal. Its purpose is to create a common foundation of tooling and practices that enables every autonomy team at Torc to experiment quickly without sacrificing rigor.  To achieve the company’s goals, Autonomy and DL+S must share a vision of the future and jointly work towards it.  Training sessions, customer discovery interviews, embedded solutions engineers, shared offsites, video happy hours, late nights in Dallas jointly debugging the double lane change… we have built a rhythm that helps us do the hard work of identifying the shared path forward and choosing to pursue it together.

I am proud of the efforts of everyone in the organization. We are working not just to improve our software but also to improve the process of building that software.  That effort will pay dividends for the immediate future and beyond, empowering us to “Drive the Future of Freight” more efficiently than ever.

Anyscale by Ray: Efficiency and Scalability of Compute

Many organizations focus on raw compute availability to scale workloads, but the real challenge is using compute effectively. Large multimodal training jobs require balancing CPU resources, GPU resources, and data throughput simultaneously. Without efficient orchestration, expensive infrastructure can sit underutilized while training workloads wait for data or processing bottlenecks to clear.

With Ray, Torc focuses on optimizing utilization rather than simply increasing capacity. shifting the focus from “Can I get more hardware?” to “Am I using the hardware effectively?”

The impact has been measurable. Recently, the team optimized a key workload for the scene modeling truck component; it provides information to the behavior models about nearby objects such as vehicles and pedestrians. They were able to improve a 4 hour training time to 1.5 hours, dropping time and costs by 73%, by improving data flow and using cheaper compute. That outcome captures the broader transformation: Ray helped move the organization from simply executing large workloads to optimally executing large workloads while improving quality.

Felix Heide presenting AV 3.0 at the 2026 Ray Summit, by Anyscale. Click below to watch his keynote address.

Torc Robotics Joins Auto-ISAC to Strengthen Cybersecurity Collaboration Across Connected Commercial Vehicles 

Torc Robotics Joins Auto-ISAC to Strengthen Cybersecurity Collaboration Across Connected Commercial Vehicles 

Torc truck on highway

Membership supports Torc’s continued focus on secure, resilient autonomous trucking systems as the company advances toward commercialization 

BLACKSBURG, Va – August 3, 2026 – Torc Robotics, a pioneer in self-driving vehicle technology, today announced it has joined the Automotive Information Sharing and Analysis Center (Auto-ISAC), a global member-driven organization dedicated to advancing cybersecurity collaboration across the automotive industry.

As autonomous trucks become more connected and integrated into the freight ecosystem, cybersecurity is an essential part of building systems that are safe, reliable and ready for commercial deployment. Through Auto-ISAC, Torc will collaborate with automakers, suppliers and commercial vehicle organizations to share cyber threat intelligence, contribute to emerging best practices and help strengthen the resilience of connected vehicle technologies.

“Cybersecurity isn’t added at the end of development; it’s built into the product from the design stage,” said Mike Westra, Senior Director of Product Cybersecurity at Torc. “As we prepare to commercialize and scale autonomous trucking, participating in Auto-ISAC helps us strengthen our approach in collaboration with the broader autonomous industry, ensuring our product is designed to address evolving cyber threats throughout its lifecycle.”

Auto-ISAC provides a trusted forum for members to share timely cyber threat intelligence, develop cybersecurity best practices and coordinate responses to emerging threats across the automotive sector. Torc’s participation reflects its broader commitment to industry collaboration as the company continues developing autonomous trucks for long-haul freight operations in the U.S.

“Delivering autonomous trucks at commercial scale requires disciplined engineering across safety, security, reliability and operations,” said Mike Avitabile, Chief Engineering Officer at Torc. “Our participation in Auto-ISAC strengthens that work by helping ensure cybersecurity advances alongside the broader technology and industry ecosystem.”

Through its membership, Torc will participate in Auto-ISAC working groups and collaborate with industry peers on cybersecurity guidance, information sharing and best practices for connected and autonomous vehicle systems.

“The cyber threat landscape is constantly evolving, and no organization can stay ahead of it alone,” said Faye Francy, executive director, Auto-ISAC. “By joining Auto-ISAC, Torc is strengthening an industry-wide community dedicated to sharing cyber threat intelligence, advancing best practices and supporting the secure deployment of autonomous commercial vehicles.”

About Torc

Torc is driving the future of freight with autonomous technology. With more than 20 years of experience in pioneering safety-critical, self-driving applications, Torc offers an AI-forward, self-driving vehicle software and integration solution and is currently focusing on commercializing autonomous trucks for long-haul applications in the U.S. In addition to its Blacksburg headquarters and engineering offices in Ann Arbor, MI, and Montreal, Torc has a fleet operations facility in Dallas-Fort Worth, to support the company’s productization and commercialization efforts for our customers. As an independent subsidiary of Daimler Truck AG, a global leader and pioneer in trucking, Torc is empowering exceptional employees, delivering a customer-focused autonomous truck product, and providing the safest, most reliable, and cost-efficient solution to the market. For more information, please visit torc.ai.

About Auto-ISAC

The Auto-ISAC was formed by automakers in 2015 to establish a global information-sharing community to address vehicle cybersecurity and operate as a central hub for sharing, tracking, and analyzing intelligence about emerging cybersecurity risks. Auto-ISAC members represent more than 99 percent of light-duty vehicles on the road in North America. Members also include heavy-duty vehicles, commercial fleets, carriers, and suppliers. For more information, please visit www.automotiveisac.com.
From Black Box to Glass Box: AV 3.0, Physical AI, and the Future of Long-Haul Trucking 

From Black Box to Glass Box: AV 3.0, Physical AI, and the Future of Long-Haul Trucking 

In 2025, Torc introduced AV 3.0, our framework for building autonomous software that is safe, scalable, and designed specifically for long-haul trucking autonomy. It powers TorcDrive, the physical Al virtual driver software that perceives, reasons, and navigates the real world in real-time.

We’ve had a lot of questions about our AV 3.0, evolution and how it differs from, and surpasses, the autonomous software industry as a whole. Here’s a breakdown of all the predecessors to our AV 3.0, what makes TorcDrive different and distinctive, and why we believe it sets the benchmark for state-of-the-art ADS systems.

Rules-Based Approaches and AV 1.0: Open Books and the First Learned Models

It all began with rules-based, open book software. In 2007, the Torc team VictorTango drove Odin, a 2005 Ford Escape hybrid designed to navigate complex urban environments without human intervention. The system successfully drove through roundabouts, intersections, and cross-traffic using a rule-based architecture, where the vehicle behavior was governed by pre-programmed, deterministic logic (if-then statements). Such early systems were highly interpretable and traceable. Each decision could be linked back to a specific rule, sensor input, or subsystem. This made it possible to diagnose failures with precision. In the mid-2010s, the emergence of deep learning enabled the initial move to AI with a massive improvement in perception performance, allowing vehicles to better detect objects, understand lane structure, and interpret complex scenes. This is what we refer to as AV 1.0, a hybrid approach combining learned perception models with rule-based prediction and planning. These architectures operated on inputs from cameras, lidar, radar, high-definition maps, and hand-coded rules of the road. However, these systems were inherently brittle, since downstream behavior prediction and planning components remained largely hand-engineered. It’s nearly impossible to imagine every scenario in advance and code appropriate rules to handle it. As a result, this architecture, combining first-generation learned perception models with rules-based prediction and planning, were not capable enough for the complexity of fully autonomous driving. Operating with distinct rules in silos, AV 1.0 vehicles were often stymied by simple situations that a human would negotiate without a second thought, like a misplaced construction cone or worn-away lane markings. AV 1.0 was far from being a scalable commercial product that could replace drivers.
Fear of a Machine You've Never Met

Most Americans are anxious about how quickly AI is advancing. That anxiety is especially pronounced with physical AI—machines that use sensors to perceive, reason, and move around in the real world in real time—like TorcDrive. But few Americans have directly interacted with any Physical AI systems.

According to polling, 87% of Americans have never ridden in an autonomous vehicle, or even know someone who has.

Not All Autonomy Is the Same

Much of society is limited to forming opinions based heavily on eye-catching headlines hyped by algorithms eager for clicks and eyeballs. All of us benefit from learning about the fundamental differences between the powerful physical AI quietly driving autonomous freight trucks and the AV systems in private vehicles and robotaxis making the evening news.

AV 2.0: A Magic Black Box

In the last five years, we’ve seen a massive leap in vehicle capabilities with AV 2.0 end-to-end learned platforms. This is one of the leading directions in AV research, moving beyond heavily rule-based planning and map-dependent architectures toward unified learned systems. Instead of relying on hand-designed rules for each part of the driving stack, AV 2.0 driving solutions leverage advances in deep learning, generative modeling, and large-scale data to train models that can run in-vehicle and connect perception, prediction, planning, and control within a single learned framework. Like the large language models we increasingly use every day, these systems are capable of impressive results. But that capability comes with a major tradeoff: opacity. In a highly integrated end-to-end model, it can be difficult to determine why the vehicle behaved a certain way, where an error originated, or how to make a targeted correction without affecting other parts of the system. A change intended to improve behavior in one scenario can introduce regressions elsewhere, requiring large-scale retraining, validation, and testing. This makes development more expensive, slows iteration, and creates uncertainty around how behavior may shift from one model update to the next. For long-haul trucking, the challenge is even greater. Safe physical AI for heavy trucks requires more than what is sufficient for robotaxis or passenger vehicles. Trucks operate with longer stopping distances, reduced maneuverability, different visibility constraints, and a need for long-range perception, prediction, and planning. The next generation of autonomous trucking must combine the deterministic confidence and traceability of rule-based autonomy, the modular structure and debuggability of AV 1.0, and the performance gains of the end-to-end approach in AV 2.0.

AV 3.0: The Transparent Glass Box

The next generation of autonomy should not force a choice between the traceability of rules-based systems and the performance of modern end-to-end AI. Torc’s approach combines modular approaches, deterministic guardrails, and end-to-end training into what we call AV 3.0: a transparent architecture for autonomous trucking. Rather than a black box approach to autonomy, AV 3.0 is designed as a “glass box”, a system whose major components can be inspected, validated, and improved with precision. This approach is structured to provide visibility into how the truck perceives the world, predicts the behavior of other road users, and selects a safe driving plan. To develop AV 3.0, Torc has the high-powered simulation and data loop needed to train, test, and validate TorcDrive before we deploy it on public roads. The result is a scalable autonomous trucking software product designed for the complexity, safety requirements, and operating realities of long-haul freight.
A graphic describing the three parts of Torc's AV 3.0

1. A Verifiable AI stack

TorcDrive automated driving software is organized around three core functional modules: Perception, Prediction, and Planning. Together, these modules allow the system to understand the driving environment, anticipate how surrounding actors may behave, and generate safe driving plans for the truck. This forms a modular learned software stack that is learned in an end-to-end manner, and is supported by deterministic, rule-based guardrails and clearly defined safety criteria. This modular structure makes the system more inspectable, testable, and verifiable than a single black-box end-to-end model. Engineers can examine intermediate outputs from each stage of the stack and validate whether the different components are responding appropriately to the input. When changes need to be made, Torc can target specific elements of the stack, validate at the module-level output, and then test the full system in closed-loop conditions before deployment. Equally important as the modular structure of the AI stack is how TorcDrive is trained and validated.

2. Immersive AI Training and Validation

Before it reaches the road, TorcDrive is trained and tested through a large-scale data and simulation loop designed to expose the system to the complexity of real-world trucking. This can be seen as the autonomous trucking equivalent of a CDL driver-education program, except the system is specifically tested in rare, complex, and safety-critical scenarios at a scale no human driver could encounter in a lifetime. Torc accomplishes this using our proprietary generative simulator trained on real-world data captured by production-intent sensor suites on over-the-road trucks. Torc’s generative AI and data-loop environment creates a large and continually expanding set of challenging driving scenarios grounded in real-world data and physics-based modeling. These scenes include cars, trucks, pedestrians, animals, varied road geometries, changing conditions, and rare edge cases that are difficult to collect repeatedly through road testing alone. Our immersive training environment accelerates TorcDrive’s ability to learn new routes and autonomous hub layouts, allows it to quickly adapt to regional driving conditions, and prepare for challenging scenarios such as severe weather, unexpected pedestrian encounters, or rare vehicle behaviors. Simulation does not replace real-world validation, but it greatly expands the range of situations TorcDrive can experience before deployment.

3. Seamless Hardware Integration

In 2019 Torc Robotics became an independent subsidiary of Daimler Truck NA, the world’s largest and leading OEM, bringing together extensive experience in freight industry manufacturing and relationships with Torc’s experience in developing autonomous vehicle solutions. This is the first strategic alliance between an autonomous vehicle technology firm and a truck original equipment manufacturer (OEM), gathering all of the essentials needed to create a scalable Class 8 SAE Level 4 truck. Together, Torc and Daimler Truck have developed a truck purpose-built for fully self-driving autonomous operation in long-haul trucking and production at scale. It combines a Verifiable AI stack hosted on a proprietary NVIDIA-powered embedded compute platform provided by Flex, with sensors, other hardware, and Daimler’s autonomous-ready 5.0 Freightliner Cascadia. This chassis is specifically designed for SAE Level 4 autonomous operation, with all the essential compute stack components and sensors installed on the production line. Complete with necessary redundancies for safety-critical components, the vehicle and platform have been proven and validated for highway operations. This is the first autonomous freight vehicle to fully transition to a production-ready, purpose-built platform—the autonomy validated from the wheels up from the beginning, not bolted-on after the fact.

Riding Along as the Future Unfolds

Over the coming weeks, we’ll be sharing more detailed pieces of our product and how it has been developed on the Al-powered technology wave. What’s down the road? Connect with us to see what’s next.

​

Digital AI vs. Physical AI: What’s the Difference and Why It Matters

Digital AI vs. Physical AI: What’s the Difference and Why It Matters

Have you used AI to streamline your work or generate a list for you? Have you been fooled by an AI image? The technology we’ve been watching in movies for 20 years is now in our homes, cars, phones, and schools. Artificial intelligence is no longer theoretical. Most people—across industries and age groups—interact with tools like ChatGPT, Claude, or Gemini. Work apps like Zoom now offer AI summaries, and Microsoft and Adobe have generative AI tools built into their products. This widespread exposure has created a shared baseline: AI is useful, real, and improving quickly.

But there’s a critical distinction that’s still not widely understood—digital AI vs. physical AI.

Digital AI:

Intelligence On-screen

Digital AI operates in a purely virtual environment. It processes text, images, and data to:
• Answer questions
• Generate content
• Analyze patterns
• Support decision-making

Its rapid rise wasn’t accidental. Digital AI benefited from a massive, ready-made dataset: human language and images. For decades, we’ve been writing, storing, and digitizing information—emails, documents, books, photos, websites. It’s been trained on the Internet: the largest repository of language and visuals in human history.

Once that data became accessible, AI systems could be trained quickly and at scale.

That’s why digital AI feels like it “appeared overnight.” It was trained “in the cloud” of ever-expanding data centers, allowing for rapid iteration.  

Physical AI:

Intelligence in the Real World

Physical AI takes the same foundational technologies as digital, on-screen AI—machine learning and neural networks—and applies them to real-world interactions.

Instead of predicting the next word, physical AI must decide:
• What object am I seeing?
• How do I move around safely?
• How hard should I grip this item?
• What action should I take next?

This introduces massive complexity. Instead of just structured data, physical AI must work within the real, unstructured environment, around people and things that don’t always follow the rules. The real world isn’t just data—it’s dynamic, unpredictable, and only governed by physics.

Physical AI needs to account for the 3D space of dimensions, with sensors, actuators, and an understanding of interactions and their outcomes.

Why Physical AI Is Slower to Scale (for Now)

When we built digital AI systems, we chose the right dataset by only using some parts of the internet. Chatbots were built using mostly social media data; Coding Agents were built using mostly open source code data.

Physical AI also requires special data for real world use cases. The world is extremely complex, and the sensors that “see” it can also be complex. But, this time, we can’t just look on the Internet to “find” it. The sensors we use, the environments in which robots move, and the special situations we have to prepare for have simply not been recorded.

Unlike digital AI, physical AI lacks a rich historical dataset like the Internet. There’s no ready made data set for how humans move through environments, how objects behave under force, or how tasks are physically completed. All the physical data must be created, not just collected.

The same generative tools that you use in digital AI can be applied within a simulated world, using real-world, grounded physics, and the physical AI hardware (robots and sensors) can be built alongside the AI. We don’t always have to wait for a robot to go out into the world, record the data, and bring it back to us – the robot isn’t ready yet! Companies like NVIDIA, Tesla, and Amazon are accelerating this by combining compute power, simulation, and real-world data collection.

At Torc, our autonomous driving system, TorcDrive, is being trained to work in the real world using both real-world data (recorded camera and lidar images of on-road driving) as well as complex simulated images created from those same on-road recordings, simultaneously.

The Key Insight: Same Brain, Different Body

The most important takeaway is this: digital AI and physical AI are built on the same core technology. Digital AI is proving what’s possible and adaptable by humans, and most importantly, helpful. Physical AI extends that capability into the real world. Everything you’ve seen AI do on a screen—learning, adapting, improving—will eventually happen in the physical world.

What This Means for the Freight Industry

For freight, logistics, and operations, digital AI helps companies and people think better … physical AI can help you perform better.

At Torc, physical AI is manifested in our autonomous driving software, TorcDrive, and powered by AV 3.0, on our trucks today, after being trained by millions of hours of real world and simulated scenarios. It’s the realization of the promises of hundreds of years of technology and human invention. We are still in early days of building physical AI but the trajectory is clear. Digital AI and physical AI aren’t separate revolutions. They are one and the same.

Meet Your First Physical AI

A physical AI machine you might be familiar with is the in-house robot vacuum. The simplest vacuum has the most basic sensors to register the world around it. If it bumps into something, the computer registers the bump against the moveable panel and turns the wheels to rotate itself in a different direction. Other more expensive models can use sensors to determine floor material, and then adjust how they clean accordingly.

More Going On Under the Hood

There are robot vacuum models with self-emptying functions, pet sensors, and even self-cleaning controls. One of the newest models introduced at CES 2025 even has a robotic arm to pick up socks. As we transition to a more physical AI world, how physical AI “understands” and interacts with its surroundings is very different “under the hood” (or dust bin, in this analogy). It’s becoming more complex, and smarter, and importantly, more applicable and helpful.

Not All Physical AI Thinks the Same

How different physical AI instances and machines “think” and how they must act on their sensors vary widely. It is critical to keep in mind that systems are running ever-increasing AI models, trained on real world and simulation data, designed to interact with the physical world in more capable ways, far exceeding the scalability (and capability) of the first physical AI technologies.

Freight Industry Physical AI Key Applications and Workflows
  • Autonomous Warehouse Operations
  • Intelligent Fleet Management & Safety
  • Dynamic Load Optimization
  • Automated Material Handling
  • And of course … Autonomous Vehicles
AV3.O
Scaling Physical AI for Autonomous Trucking Today: Torc’s First-to-Production Embedded Hardware

Scaling Physical AI for Autonomous Trucking Today: Torc’s First-to-Production Embedded Hardware

Flex Jupiter unit for Torc

In the first article about Torc’s AV 3.0 technology, we covered the different system components of AV 3.0: the virtual driver software, and the advanced data loop and generative AI simulation infrastructure to train and test it. AV 3.0 is the robust technology needed to safely and efficiently create an autonomous trucking product.

However, AV 3.0 is more than just world class autonomous software. We’re running this physical AI product on high-volume embedded compute in production-ready vehicles today.

We’re the first in the industry to deploy our system on a production-intent platform — Daimler Truck’s autonomous-ready Freightliner Cascadia. By running our AV 3.0 system on this autonomy-specific redundant chassis, Torc is setting the industry standard in the autonomous trucking space and leading the way to commercially scalable and viable freight solutions.

A graphic describing the three parts of Torc's AV 3.0

The Embedded Hardware Engine of AV 3.0

Freight movement presents unique challenges that require unique trucking solutions. Systems like long range perception, long-distance-actor prediction, and extended behavior planning are non-negotiable. Add in that the physical compute platform needs to hold up to running millions of miles in harsh long-haul trucking environments.

We worked with Flex to optimize the performance, cost, power, and reliability of our embedded compute platform to meet the stringent requirements needed for running Torc’s AV 3.0 virtual driver. This is no commercial server rack setup packed full of GPUs in the cab.

 It’s optimized to drive the future of autonomous freight and nothing else.

Our multi-chip adaptable architecture leverages:

  • Flex’s Jupiter platform
  • NVIDIA DRIVE AGX
  • NVIDIA DRIVE Orin system-on-a-chip
  • NVIDIA DriveOS operating system

Rishi Dhall, Vice President of Automotive at NVIDIA (when the partnership was announced in March 2025 ) said, “NVIDIA DRIVE AGX has been industry-proven in full production for automotive real-time applications at the edge. It delivers the high compute performance, low latency, and multi-sensor connectivity needed for Torc’s sophisticated autonomous trucking software, delivering robust perception, prediction, and planning for safe and reliable operation. Torc is on a clear path to scalable production for its commercial launch in 2027.”

Mike Thoeny, President of automotive at Flex, was also quoted at the time: “Our collaboration with Torc, Daimler Truck, and NVIDIA illustrates how Flex partners across the full ecosystem to enable mobility companies to launch next-generation technology with greater resilience and speed. We appreciate the trust Torc and Daimler Truck have placed in Flex through leveraging our Jupiter compute platform and advanced manufacturing capabilities to deliver autonomous long-haul trucking at scale.”

Torc’s work to achieve AV 3.0 has been intense, and the timeframe impossibly short, but this necessary step was anticipated years ago and well executed by the hundreds of dedicated Torc’rs that made it possible. Torc has staked our claim amongst the competition as the first autonomous freight company to fully shift to a production embedded platform.

FLEX unit
MarsII unit by Flex for Level 2
Advanced Driver-Assistance System (ADAS) Applications.
Image courtesy of Flex.
Jupiter-T unit for Torc. Image courtesy of Flex

Jupiter-T unit for Torc. Level 4 ADAS. Image courtesy of Flex

“The incredibly intensive, time-consuming, and technically challenging effort that it took to take the virtual driver software that runs on ‘unlimited’ datacenter server racks and make it work on the highly reliable Flex embedded compute platform, all while not sacrificing performance, is a technological marvel.”

– Stephan Vargas, Torc Technology Vice President (Compute Foundation)

Pieces of the Whole

In 2019, Torc made history by partnering with Daimler Truck, creating the first OEM relationship in the business. Torc’s leaders knew this partnership was the only path to delivering a safe and scalable product on time. Both companies knew the Cascadia Freightliner would be the top choice to house the industry’s best-in-class autonomous truck – and it’s all been realized.

Follow along as we map out what happened over the last year and why it’s important now.

October 2024: Torc Robotics Performs Successful Fully Autonomous Product Validation

In October 2024, the driverless production-intent truck, redundant components, AD Kit, and the first release of the production-intent virtual driver took to the road on a closed multi-lane test track. Clocking in at over five hours in highway settings, the test was conducted at full operating speeds of up to 65 mph to optimize fuel efficiency and emulate commercial timings and conditions for future long-haul routes. The test signaled a sea-change for Torc, shifting cycles from development to productization. This occasion was no demonstration. It was the proof point for years of work. The fully robotic from hub-to-surface-to-ramp-to-highway and back production intent hardware and software were successfully on the road.

March 2025: Torc Collaborates with Flex on Physical AI Platform for Autonomous Trucks, Accelerated by NVIDIA

In March 2025, we attended GTC in San Jose, where we announced our collaboration with NVIDIA and Flex.

April 2025: Daimler Truck’s Autonomous-Ready Fifth Generation Freightliner Cascadia Hits Texas Roads With Torc

To enable SAE Level 4 autonomous driving, the company has purposefully designed and built redundancy into the Freightliner Cascadia platform for safety-critical systems for safe, driverless operations. With over 1,500 engineering requirements, all translated into features, and a second set of electronically controlled systems like an integrated power network, the autonomous-ready Cascadia sets an industry standard for autonomous systems integration. – Daimler Truck press release April 2025

The autonomous-ready Freightliner Cascadia announced by Daimler Truck in April 2025 includes all essential compute stack components and sensors installed on the production line; a step that allows seamless integrate of the virtual driver software without retrofitting the platform.

Importantly, the vehicle and platform had already been proven and validated at our driverless product acceptance test in October 2024, six months prior.

 

May 2025: Torc Commercial Center Operations Commence

After 20 years in the robotics business, and six years of software research and development, Torc officially opened the company’s first commercial hub in May 2025, getting our trucks officially on the I-35 corridor. Our fleet has moved from our testing grounds in Albuquerque to focus on the Laredo to Fort Worth route, a crucial freight lane for many large freight customers. We’re already working in concert with them as we continue to look for more opportunities for the latter half of 2026.

 

August 2025: Torc Technology Center Opens

Sunny drone shot of the Torc Ann Arbor office locationThere are still technology refinements and milestones to hit for our virtual driver and AI development as well, and we’ll be doing that with our new technology center in Ann Arbor, Michigan. Situated close to closed course test tracks and surrounded by top-notch robotics university programs, we continue to onboard the Torc’rs needed to get to our market entry timeframe.

 

September 2025: AV 3.0 Technology Debuts

Sunny drone shot of the Torc Ann Arbor office locationThere are still technology refinements and milestones to hit for our virtual driver and AI development as well, and we’ll be doing that with our new technology center in Ann Arbor, Michigan. Situated close to closed course test tracks and surrounded by top-notch robotics university programs, we continue to onboard the Torc’rs needed to get to our market entry timeframe.

 

“We’ve never been focused on chasing timelines, or on demonstrations or headlines. We’ve just been focused on a safe and scalable product from the beginning and just doing what’s right for the industry and our partners.”

– Peter Vaughn Schmidt, CEO

Onward To 2027

We’re leading the wave of physical AI applications with our simulation environment, our computation and sensing hardware are defining the industry, and future customers are getting ride-alongs inside our production-intent trucks in Texas today. We’re set up to immediately scale and deliver on customer demand.

So, now what? What will we be doing until market entry?

Over the coming months, we’ll be sharing more about hub operations and the development of workflows and job creation, customer autonomous fleet onboarding transition plans, our commitment to safety and the release of our safety case, and more cutting-edge simulations and visualizations from our technology and trucks.  We have time to solve a multitude of other necessary autonomous trucking puzzles, truly making Torc the industry leader, driving the future of freight.

Follow us on our social channels to keep up to date with all our announcements and upcoming events.

“There are two kinds of autonomous trucks on the road today, at the end of 2025. There are demonstration trucks – and there are Torc trucks.”

– Andrew Culhane, Chief Commercial Officer