Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Perspective-view self-play can train a camera-based driving policy directly, without privileged observations, and transfers zero-shot to new layouts better than its vectorized counterpart.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 00:53 UTC pith:TJIK2KKU

load-bearing objection A genuinely new simulator-plus-training result that makes image-based self-play practical at scale, but the claim that the privileged/distillation stage is unnecessary is not actually tested against a distilled camera-student baseline. the 3 major comments →

arxiv 2607.26005 v1 pith:TJIK2KKU submitted 2026-07-28 cs.CV cs.AIcs.RO

Pictura: Perspective-View Self-Play at Scale for Driving

Yuan Yin , Elias Ramzi , Marc Lafon , Valentin Charraut , Victor Bares , Yihong Xu , \'Eloi Zablocki , Alexandre Boulch
show 3 more authors
Thibault Buhet Andrei Bursuc Matthieu Cord
This is my paper
classification cs.CV cs.AIcs.RO
keywords self-playautonomous drivingperspective viewreinforcement learningrasterizationegocentric perceptionzero-shot transferend-to-end driving
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that the standard recipe for large-scale driving self-play — train on privileged vectorized state, then distill into a camera-based student — is unnecessary. It introduces a simulator whose renderer draws each agent's egocentric perspective view cheaply enough to sit inside the reinforcement-learning loop, and trains a policy from scratch on those rendered images alone. That policy nearly matches a privileged vectorized counterpart on training-domain scenarios and outperforms it zero-shot on real-world logged layouts re-rendered through the same simulator. If correct, the representation gap between simulation and deployment is addressed at its source rather than patched by imitation of decisions a camera cannot justify.

Core claim

The discovery is that self-play at 50 billion agent steps can run directly on perspective images, with no privileged observations and no distillation stage. The resulting policy, Alberti, learns to drive from what its cameras see, develops occlusion-aware caution on its own, and transfers zero-shot to layouts derived from real-world driving logs more safely than privileged vectorized agents. The paper explicitly frames alignment with real camera imagery as a separate future step; its zero-shot evidence comes from re-rendering logged layouts through the same rasterizer.

What carries the argument

The load-bearing mechanism is Pictura's GPU rasterizer, which renders each agent's four-camera egocentric view as flat-shaded geometric primitives on the compute cores inside the training loop, sustaining about 500K agent-steps per second (2M images/s). This makes the perspective view cheap enough to be the RL observation itself, so perception and control are learned jointly under the constraint that the policy can only act on what a camera could see. The policy pools rendered views with per-camera cross-attention queries before a shared actor-critic trunk trained with standard reinforcement learning.

Load-bearing premise

The claim rests on the premise that a flat-shaded rasterized projection of simulator state is equivalent, for learning and transfer, to what a real camera image would show; if that visual grounding does not carry over to real frames, the deployment value is unproven.

What would settle it

Freeze Alberti and run it on real camera frames, or photorealistic renders, of the same road layouts without fine-tuning or domain adaptation. If collision and goal-completion metrics collapse relative to the re-rendered-layout evaluation while a same-architecture policy trained on those real frames succeeds, the central claim that perspective-view self-play closes the representation gap is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Privileged observations and student distillation are not required for self-play driving; a camera-input policy can be trained directly at scale.
  • A policy trained only on rendered perspective views slows at occluded corners and where its sight ends, re-accelerating when the road is clear — behavior that emerges without explicit occlusion reward shaping.
  • Zero-shot transfer to real-world logged layouts improves as the rasterizer maps different map sources into a common visual vocabulary, whereas vectorized observations expose metric distributions that shift between map sources.
  • Render resolution buys goal completion and distance per infraction at the cost of throughput; a resolution curriculum late in training should capture most of the gain cheaply.
  • Because rendering occupies only about 10% of training time on the compute-core rasterizer, further scaling of image-based self-play is not bottlenecked by rendering.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the red-light violation gap suggests traffic-light resolution is the next bottleneck; a higher-resolution crop around signals or a dedicated signal detector could close it without full-image cost.
  • Editorial inference: the zero-shot result is about the rasterizer harmonizing map sources, not about real camera images; whether the learned grounding survives the jump to real frames remains the decisive test.
  • Editorial inference: the counterfactual deletion probes imply a natural evaluation protocol — measuring whether a policy's value and braking responses vanish for fully occluded agents — that could be adopted by other sim-to-real pipelines.
  • Editorial inference: because the policy is trained from scratch on cheap renders, the same pipeline could be pointed at other sensor views, such as a single wide camera or fisheye, to test how much of the benefit comes from field-of-view coverage versus perspective projection itself.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents Pictura, a GPU-accelerated multi-agent driving simulator with a custom CUDA rasterizer that renders each agent's egocentric perspective view at every simulation step at high throughput (500K agent-steps/s, 2M images/s on one H100). Using Pictura, the authors train Alberti, a from-scratch PPO policy that observes four rendered camera views and no privileged vectorized state, over 50B agent steps. They compare against privileged vectorized baselines in-domain on CARLA maps and zero-shot on WOMD layouts re-rendered through Pictura. Alberti approaches the vectorized baselines on off-road and goal metrics in-domain, but has higher collision and red-light violation rates; on zero-shot WOMD it outperforms both vectorized baselines. The paper also analyzes visual grounding via counterfactual probes and closed-loop blind-corner takeovers, and provides detailed renderer efficiency, memory, and wall-clock ablations.

Significance. If the central claim holds, the paper could remove a common design requirement—privileged vectorized self-play followed by distillation—from large-scale driving RL. The engineering contributions are substantial: a compute-core rasterizer integrated into the RL loop, throughput far above prior perspective simulators, careful ablation of resolution and wall-clock trade-offs, and a large-scale 50B-step demonstration. The counterfactual and blind-corner analyses are a valuable step toward understanding what image-based policies ground on. However, the significance hinges on two unverified premises: that a distilled camera-student would not beat direct image-RL, and that the rendered perspective view is representative enough of real camera images to support deployment. Both are acknowledged or omitted rather than resolved.

major comments (3)
  1. [Sec. 1 / Tab. 1a] The central claim that the privileged/distillation stage is 'unnecessary' (Sec. 1, para. 4) is not tested against the actual alternative. All comparisons in Tab. 1 are against vectorized privileged agents (Vectorized, Vectorized†); no camera-input student distilled from either baseline is trained or evaluated. Sec. 5.2 argues qualitatively (citing LEAD) that distillation inherits non-causal behavior, but no distillation experiment is reported. Without such a baseline, the paper establishes that direct image-RL is feasible and competitive, but cannot exclude that a distilled student would achieve better safety metrics (especially collision and red-light) on the same observations. Please add a distilled camera-student baseline (e.g., BC from the privileged teacher or RL with distilled rewards) or soften the 'unnecessary' claim.
  2. [Sec. 1 / Sec. C.2] The zero-shot claim is limited by two factors. First, WOMD layouts are re-rendered through Pictura's rasterizer; no real camera images are used, and Sec. 1 defers alignment to real imagery as 'a complementary, orthogonal step.' Thus the experiments demonstrate transfer to new maps/agents in the same rendered observation space, not transfer to real camera observations. Second, Sec. C.2 reports that the vectorized baseline's polyline subdivision span was chosen by a sweep over the zero-shot test set ('A sweep over that span puts the best zero-shot performance at 5m'), i.e., test-set selection. This is not a strict zero-shot evaluation for the vectorized baseline. While the selection favors the baseline, the comparison is not the result of a fixed protocol, and sensitivity to this hyperparameter is not reported. Please state these limitations explicitly or scope the claim to the rendered do
  3. [Sec. 5.1 / Tab. 1a] In-domain, the 'approaches' claim is only partially supported. At medium density, Alberti has collision 0.060 vs 0.037 (Vectorized†) and red-light 0.058 vs 0.009; km/infr. is 11.61 vs 26.99. These are safety-critical gaps, not minor. The explanation (small traffic-light pixels) is plausible but does not change the fact that a practical camera-based policy is not yet demonstrated in-domain. Please either improve these metrics (e.g., resolution/architecture/augmentation targeted at small objects) or report the claim as 'approaches on off-road and goal metrics, with substantial gaps in collision and red-light.'
minor comments (5)
  1. [Sec. 3.2 / Eq. (2)] The sentence 'the ego state S_i and the conditioning C_i stay vector inputs, matching what is observable on-board at test time' is questionable: C_i contains reward and dynamics sampling coefficients (Tab. 2) that are not on-board observable. Please clarify how C_i is set at deployment.
  2. [Sec. 5.1 / Sec. A.3] The paper says Alberti uses 'the same training recipe and hyperparameters as the vectorized agent,' but Alberti was trained on 32 H100s while vectorized baselines used 4 H100s. This is not a problem for the transition-matched comparison, but it should be stated that compute differs.
  3. [Fig. 5a] The throughput comparison pulls RAP and HUGSIM numbers from [25]; consider adding error bars or confidence intervals since these are single reported values.
  4. [Sec. 5.2 / Fig. 8] Please describe how the 16 scenes and 88K removals were sampled; reporting standard deviations across scenes would strengthen the aggregate result.
  5. [Sec. 1 / Abstract] The abstract says 'first large-scale driving self-play policy trained directly from perspective images' and the conclusion says 'to our knowledge, the first self-play driving policy trained from rendered camera views.' Please verify this claim against the cited related work (e.g., Gigapixel, RAP) and soften if needed.

Circularity Check

0 steps flagged

No significant circularity; central claims are anchored by controlled baselines and an external WOMD benchmark.

full rationale

This is an empirical systems paper, not a derivation, and I find no step where a prediction reduces to a fitted input or to a self-citation chain. Alberti is trained by plain PPO on rendered observations in CARLA-style maps; the headline claims are measured either in-domain on newly generated scenarios or zero-shot on WOMD layouts. The zero-shot protocol re-renders WOMD through Pictura, but this is a deliberate evaluation design for testing map/agent-state generalization, not a way of injecting WOMD statistics into training: no WOMD data is used during training, and the rasterizer is a fixed deterministic function rather than a fitted model. The self-citations (RAP [11] for visual design, DrivoR [16] for architecture, Gigaflow [8] for reward/recipe) support implementation choices and training conventions, not the central claim that perspective-view self-play is practical. The central comparisons are against vectorized baselines trained under the same recipe, including a like-for-like variant that removes privileged auxiliary features. Section 5.2's grounding analysis is an empirical probe of the trained policy's responses to visible versus occluded agents, not a derivation of the training outcome from its own assumptions. The acknowledged limitation that real-camera alignment is left to future work is a scope boundary, not a circular step. No fitted parameter is renamed as a prediction, and no uniqueness theorem or self-citation is invoked to force a conclusion. The paper is self-contained against external benchmarks, so it receives a 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The ledger for this systems paper is dominated by hand-chosen training/evaluation hyperparameters and domain assumptions rather than fitted physical constants. The most consequential are the reward/dynamics conditioning distributions, the default render resolution/camera rig, and the test-set-tuned WOMD polyline subdivision span for the vectorized baseline. No invented entities appear.

free parameters (5)
  • Reward and dynamics conditioning distributions (C_reward, C_dynamics) = Ranges in Tab. 2, e.g., α_collision ~ U(0,3), α_off-road ~ U(0,3), C_throttle ~ Umix(0.8,1.25)
    Resampled per agent per episode and fed to the policy through C_i; these hand-chosen shaping terms (inherited/modified from Gigaflow [8]) directly determine the behaviors learned, and the paper gives no sensitivity analysis.
  • PPO/GAE recipe hyperparameters = Tab. 5: clip 0.2, entropy 0.01, advantage filter 0.01×running max, rollout 128, minibatch 65,536, LR 5e-4 cosine
    Fixed following Gigaflow/PPO defaults; the central 'no perception-specific tuning' claim refers to this recipe, but no ablation shows the results are robust to these choices.
  • Default render resolution and camera rig = 96×54 per camera, 4 cameras (Tab. 3)
    Resolution sweep (Fig. 7) shows goal completion and km/infr. improve with resolution, so the default trades performance for throughput; a design choice affecting all headline numbers.
  • WOMD polyline subdivision span (vectorized baseline) = 5 m
    Chosen by sweeping on the zero-shot WOMD evaluation set (Sec. C.2); test-set hyperparameter selection, which improves the vectorized baseline rather than Alberti.
  • Alberti architecture dimensions = Tab. 4: 5-layer ConvNet (128 ch), 16 learned queries/camera, 4×512 MLP backbone
    Hand-designed token-pooling architecture; not varied in the main results, so all metrics depend on these sizes.
axioms (5)
  • domain assumption Rasterized non-photorealistic perspective views carry sufficient information for learning driving and for later transfer to real images after alignment.
    Sec. 1: 'Growing evidence indicates that scene structure and dynamics, more than appearance, carry the learning signal...' The paper relies on this to justify dropping appearance; if it fails, the sim-to-real value is unsupported.
  • domain assumption The conditioning vector C_i (reward weights and dynamics multipliers) is available at test time and is not a privileged observation.
    Eq. 2 and Sec. 3.2 keep C_i as vector inputs, described as 'matching what is observable on-board at test time'; this is asserted, not proven.
  • domain assumption Self-play against copies of itself at sufficient scale yields robust, transferable driving behavior.
    Adopted wholesale from Gigaflow [8] as the starting premise (Sec. 1 and 3.1); not re-derived.
  • domain assumption The hand-defined infraction metrics (at-fault collision, off-road, red-light, goals) are a valid proxy for driving quality.
    Sec. C.1 defines these; no correlation with real-world safety is provided.
  • domain assumption Re-rendering WOMD layouts through Pictura constitutes a valid zero-shot distribution shift for driving policies.
    Sec. 5.1: the rasterizer maps CARLA and WOMD into the same visual vocabulary, reducing the observation shift for Alberti; the generality of this transfer is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 20507 in / 21961 out tokens · 206686 ms · 2026-08-01T00:53:26.240661+00:00 · methodology

0 comments
read the original abstract

Self-play in simulation produces robust driving policies at scale. Demonstrations of such behavior have been made using privileged vectorized observations such as exact poses and velocities, even for occluded agents. This assumes that perception is solved and introduces a representation gap with the partial observation of a deployed agent driving from the perspective view of egocentric cameras. A common fix, distilling the privileged policy into a camera-input student, leaves the student imitating decisions its own view cannot justify. Instead, we establish perspective-view self-play as a practical training regime. We introduce Pictura, a GPU-accelerated multi-agent driving simulator that renders each agent's egocentric view at every step, mitigating the representation gap at its source. Pictura sustains up to 500K agent-steps/s (2M images/s) on a single H100. Using Pictura, we train Alberti by self-play with plain PPO. It is the first large-scale driving self-play policy trained directly from perspective images, without privileged observations. Training spans 50B agent steps for ~35M km of driving. It approaches the driving performance of its privileged vectorized counterpart, and transfers zero-shot to Waymo Open Motion Dataset layouts re-rendered in Pictura, where it outperforms privileged vectorized agents. Project page: https://valeoai.github.io/Pictura/

Figures

Figures reproduced from arXiv: 2607.26005 by Alexandre Boulch, Andrei Bursuc, Elias Ramzi, \'Eloi Zablocki, Marc Lafon, Matthieu Cord, Thibault Buhet, Valentin Charraut, Victor Bares, Yihong Xu, Yuan Yin.

Figure 1
Figure 1. Figure 1: Pictura runs driving self-play directly in perspective view. Left: a rendered frame of the kind the policy learns from, produced by our custom rasterizer; the labels mark the scene elements it supports beyond moving vehicles. Right: the scale of the longest self-play run of Alberti. driving behavior emerges from self-play at scale, with no human demonstrations, transferring zero-shot to maps and agents uns… view at source ↗
Figure 2
Figure 2. Figure 2: Pictura architecture. Pictura couples a vectorized simulator with a custom GPU rasterizer that renders the simulator state into each agent’s egocentric perspective view. The observation feeds Alberti, the policy, and, together with the action and the reward, forms the rollout that PPO turns into a policy update. the perspective rasterizer (Sec. 3.2), and the training recipe and policy behind Alberti (Sec. … view at source ↗
Figure 3
Figure 3. Figure 3: Rendering features for driving. Elements Pictura draws into each agent’s egocentric view, grouped by type. Agents are cuboids colored per face with depth-aware brightness: pedestrians (a), cyclists (b), vehicles (c). Traffic lights are disks with limited angular visibility, drawn only to the approach they govern (d) and culled once the camera faces away (e); the far-side heads of American-style intersectio… view at source ↗
Figure 4
Figure 4. Figure 4: Alberti architecture. A from￾scratch ConvNet encodes each cam￾era view, and per-camera query tokens cross-attend over its feature map to pool it into scene tokens. The rest is unchanged from the vectorized agent. Policy architecture. Inspired by DrivoR [16], the perspective policy encodes each cam￾era view I (t) i,k of Eq. 2 with a small convo￾lutional stack, trained from scratch, and pools the resulting f… view at source ↗
Figure 5
Figure 5. Figure 5: Pictura renderer performance. (a) Render-only agent-steps/second vs. resolution. Left: all systems on the same A100L; ⋆numbers from [25]. Right: hardware scaling of the Pictura CUDA renderer vs. its Madrona reimplementation, on A100L and H100. (b) Per-phase wall-clock of a training iteration through render resolutions on a single H100 NVL, absolute seconds (left) and share of the iteration (right). rendere… view at source ↗
Figure 6
Figure 6. Figure 6: In-domain driving performance of Alberti over 50 B steps of self-play. The bold curve is medium traffic density; the shaded band spans light to dense. vs. 12.81), while exhibiting a moderately higher collision rate (0.060 vs. 0.037). Its clearest gap lies in red-light violations (0.058 vs. 0.009), which also largely account for its lower distance per infraction (11.61 vs. 26.99 km). This gap may arise beca… view at source ↗
Figure 7
Figure 7. Figure 7: Rendering resolution sweep. In-domain self-play driving metrics and training throughput of 10 B-step Alberti agents across render resolutions; a policy is trained for each resolution. The dot is the medium traffic density; the vertical lines spans the range across densities. Rightmost: peak self-play training throughput on 32 H100s. staying the safest and reaching the most goals at every level (100%, 75%, … view at source ↗
Figure 8
Figure 8. Figure 8: Response of value and braking probability to a single nearby agent. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Closed-loop behavior at blind corners. For the same recorded state, Alberti (blue) and the privileged vectorized agent (red) take over the ego vehicle, while all other agents follow their logged trajectories. (a) Mean ego speed with an occluded oncoming vehicle (left) and at comparable empty blind corners (right); the gray band marks the main reveal times. (b) Example takeover with cross traffic occluded b… view at source ↗
Figure 10
Figure 10. Figure 10: Rendering across resolutions: CUDA vs. Madrona. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Training-step wall-clock breakdown including the quality-matched [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Wall-clock comparison: Pictura (CUDA rasterizer) vs. the Madrona reimplementation. The same perspective agent trained with either renderer at 64×36, both to 2 B agent steps (Pictura (CUDA), solid blue; Madrona, dashed red). Each panel tracks one vehicle-only metric, evaluated on held-out gigaflow self-play at medium density (120 agents, bold line; band spans light 60 to dense 150), against the wall-clock … view at source ↗
Figure 13
Figure 13. Figure 13: Vectorized self-play over training. Self-play metrics of the two 50 B vectorized policies across training, with and without the auxiliary features a perspective agent cannot perceive. Medium bold; band spans light–dense; the faint blue dashed curve is Alberti’s medium-density trajectory ( [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 7 linked inside Pith

  1. [1]

    In: ICRA (2022) 4

    Amini, A., Wang, T., Gilitschenski, I., Schwarting, W., Liu, Z., Han, S., Karaman, S., Rus, D.: VISTA 2.0: An open, data-driven simulator for multimodal sensing and policy learning for autonomous vehicles. In: ICRA (2022) 4

    2022

  2. [2]

    arXiv preprint arXiv:2106.11810 (2021) 4, 19, 20

    Caesar, H., Kabzan, J., Tan, K.S., Fong, W.K., Wolff, E.M., Lang, A.H., Fletcher, L., Beijbom, O., Omari, S.: nuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810 (2021) 4, 19, 20

    Pith/arXiv arXiv 2021

  3. [3]

    In: ICCV (2021) 4

    Chen, D., Koltun, V., Krähenbühl, P.: Learning to drive from a world on rails. In: ICCV (2021) 4

    2021

  4. [4]

    In: CoRL (2019) 4

    Chen, D., Zhou, B., Koltun, V., Krähenbühl, P.: Learning by cheating. In: CoRL (2019) 4

    2019

  5. [5]

    com/Emerge-Lab/PufferDrive2, 4, 5

    Cornelisse, D., Cheng, S., Mandavilli, P., Hunt, J., Joseph, K., Doulazmi, W., Charraut, V., Gupta, A., Suarez, J., Vinitsky, E.: PufferDrive: A fast and friendly driving simulator for training and evaluating RL agents (2025),https://github. com/Emerge-Lab/PufferDrive2, 4, 5

    2025

  6. [6]

    arXiv preprint arXiv:2606.19370 (2026) 3

    Cornelisse, D., Hunt, J., Zhang, Z., Doulazmi, W., Joseph, K., Fisac, J.F., Vinitsky, E.: Human-like autonomy emerges from self-play and a pinch of human data. arXiv preprint arXiv:2606.19370 (2026) 3

    Pith/arXiv arXiv 2026

  7. [7]

    Reinforcement Learning Journal1 (2024) 3

    Cornelisse, D., Vinitsky, E.: Human-compatible driving agents through data- regularized self-play reinforcement learning. Reinforcement Learning Journal1 (2024) 3

    2024

  8. [8]

    In: ICML (2025) 2, 3, 5, 8, 18, 20

    Cusumano-Towner, M.F., Hafner, D., Hertzberg, A., Huval, B., Petrenko, A., Vinitsky, E., Wijmans, E., Killian, T.W., Bowers, S., Sener, O., Krähenbühl, P., Koltun, V.: Robust autonomy emerges from self-play. In: ICML (2025) 2, 3, 5, 8, 18, 20

    2025

  9. [9]

    In: CoRL (2017) 3, 4, 5

    Dosovitskiy, A., Ros, G., Codevilla, F., López, A.M., Koltun, V.: CARLA: an open urban driving simulator. In: CoRL (2017) 3, 4, 5

    2017

  10. [10]

    In: ICCV (2021) 3, 4, 10, 26

    Ettinger, S., Cheng, S., Caine, B., Liu, C., Zhao, H., Pradhan, S., Chai, Y., Sapp, B., Qi, C.R., Zhou, Y., Yang, Z., Chouard, A., Sun, P., Ngiam, J., Vasudevan, V., McCauley, A., Shlens, J., Anguelov, D.: Large scale interactive motion forecasting for autonomous driving : The waymo open motion dataset. In: ICCV (2021) 3, 4, 10, 26

    2021

  11. [11]

    In: ICLR (2026) 3, 4, 6, 8, 14 16 Y

    Feng, L., Gao, Y., Zablocki, É., Li, Q., Li, W., Liu, S., Cord, M., Alahi, A.: RAP: 3d rasterization augmented end-to-end planning. In: ICLR (2026) 3, 4, 6, 8, 14 16 Y. Yin et al

    2026

  12. [12]

    In: NeurIPS Datasets and Benchmarks (2023) 4

    Gulino, C., Fu, J., Luo, W., Tucker, G., Bronstein, E., Lu, Y., Harb, J., Pan, X., Wang, Y., Chen, X., Co-Reyes, J.D., Agarwal, R., Roelofs, R., Lu, Y., Montali, N., Mougin, P., Yang, Z., White, B., Faust, A., McAllister, R., Anguelov, D., Sapp, B.: Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In: NeurIPS Data...

    2023

  13. [13]

    In: CVPR (2023) 1

    Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., Lu, L., Jia, X., Liu, Q., Dai, J., Qiao, Y., Li, H.: Planning-oriented autonomous driving. In: CVPR (2023) 1

    2023

  14. [14]

    In: ICCV (2023) 4

    Jia, X., Gao, Y., Chen, L., Yan, J., Liu, P.L., Li, H.: DriveAdapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In: ICCV (2023) 4

    2023

  15. [15]

    In: ICLR (2025) 2, 4

    Kazemkhani, S., Pandya, A., Cornelisse, D., Shacklett, B., Vinitsky, E.: Gpudrive: Data-driven, multi-agent driving simulation at 1 million FPS. In: ICLR (2025) 2, 4

    2025

  16. [16]

    In: CVPR (2026) 1, 4, 8

    Kirby, E., Boulch, A., Xu, Y., Yin, Y., Puy, G., Zablocki, É., Bursuc, A., Gidaris, S., Marlet, R., Bartoccioni, F., Cao, A.Q., Samet, N., Vu, T.H., Cord, M.: Driving on registers. In: CVPR (2026) 1, 4, 8

    2026

  17. [17]

    In: NeurIPS (2017) 3

    Lanctot, M., Zambaldi, V., Gruslys, A., Lazaridou, A., Tuyls, K., Pérolat, J., Silver, D., Graepel, T.: A unified game-theoretic approach to multiagent reinforcement learning. In: NeurIPS (2017) 3

    2017

  18. [18]

    IEEE Trans

    Li, Q., Peng, Z., Feng, L., Zhang, Q., Xue, Z., Zhou, B.: MetaDrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE Trans. Pattern Anal. Mach. Intell.45(3), 3461–3475 (2023) 4

    2023

  19. [19]

    In: ECCV (2024) 4

    Li, X., Zhang, Y., Ye, X.: DrivingDiffusion: Layout-guided multi-view driving scenarios video generation with latent diffusion model. In: ECCV (2024) 4

    2024

  20. [20]

    In: ECCV (2024) 4

    Ljungbergh, W., Tonderski, A., Johnander, J., Caesar, H., Åström, K., Felsberg, M., Petersson, C.: NeuroNCAP: Photorealistic closed-loop safety testing for autonomous driving. In: ECCV (2024) 4

    2024

  21. [21]

    In: CVPR (2026) 2, 4, 14

    Nguyen, L., Fauth, M., Jaeger, B., Dauner, D., Igl, M., Geiger, A., Chitta, K.: Lead: Minimizing learner-expert asymmetry in end-to-end driving. In: CVPR (2026) 2, 4, 14

    2026

  22. [22]

    In: NIPS (1988) 1

    Pomerleau, D.A.: Alvinn: An autonomous land vehicle in a neural network. In: NIPS (1988) 1

    1988

  23. [23]

    In: CVPR (2026) 3

    Rahimi, A., Gerard, V., Zablocki, É., Cord, M., Alahi, A.: MAD: motion appearance decoupling for efficient driving world models. In: CVPR (2026) 3

    2026

  24. [24]

    arXiv preprint arXiv:2506.09042 (2025) 4

    Ren, X., Lu, Y., Cao, T., Gao, R., Huang, S., Sabour, A., Shen, T., Pfaff, T., Wu, J.Z., Chen, R., Kim, S.W., Gao, J., Leal-Taixé, L., Chen, M., Fidler, S., Ling, H.: Cosmos-drive-dreams: Scalable synthetic driving data generation with world foundation models. arXiv preprint arXiv:2506.09042 (2025) 4

    Pith/arXiv arXiv 2025

  25. [25]

    arXiv preprint arXiv:2606.19641 (2026) 2, 3, 4, 8, 9, 14

    Rowe, L., Girgis, R., de Schaetzen, R., Cornelisse, D., Grandhi, A., Heide, F., Vinitsky, E., Pal, C., Paull, L.: Scaling self-play for end-to-end driving. arXiv preprint arXiv:2606.19641 (2026) 2, 3, 4, 8, 9, 14

    Pith/arXiv arXiv 2026

  26. [26]

    In: ICLR (2016) 8

    Schulman, J., Moritz, P., Levine, S., Jordan, M.I., Abbeel, P.: High-dimensional continuous control using generalized advantage estimation. In: ICLR (2016) 8

    2016

  27. [27]

    arXiv preprint arXiv:1707.06347 (2017) 8

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017) 8

    Pith/arXiv arXiv 2017

  28. [28]

    ACM Transactions on Graphics (2023) 4, 9 Pictura: Perspective Self-Play for Driving 17

    Shacklett, B., Rosenzweig, L.G., Xie, Z., Sarkar, B., Szot, A., Wijmans, E., Koltun, V., Batra, D., Fatahalian, K.: An extensible, data-oriented architecture for high- performance, many-world simulation. ACM Transactions on Graphics (2023) 4, 9 Pictura: Perspective Self-Play for Driving 17

    2023

  29. [29]

    Science (2018) 3

    Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., Lillicrap, T.P., Simonyan, K., Hassabis, D.: A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science (2018) 3

    2018

  30. [30]

    arXiv preprint arXiv:2406.12905 (2024) 5

    Suarez, J.: PufferLib: Making reinforcement learning libraries and environments play nice. arXiv preprint arXiv:2406.12905 (2024) 5

    Pith/arXiv arXiv 2024

  31. [31]

    In: ICCV Au- tonomous Driving Workshop (2019) 3

    Tang, Y.: Towards learning multi-agent negotiations via self-play. In: ICCV Au- tonomous Driving Workshop (2019) 3

    2019

  32. [32]

    Communications of the ACM (1995) 3

    Tesauro, G.: Temporal difference learning and td-gammon. Communications of the ACM (1995) 3

    1995

  33. [33]

    In: CVPR (2024) 4

    Wen, Y., Zhao, Y., Liu, Y., Jia, F., Wang, Y., Luo, C., Zhang, C., Wang, T., Sun, X., Zhang, X.: Panacea: Panoramic and controllable video generation for autonomous driving. In: CVPR (2024) 4

    2024

  34. [34]

    arXiv preprint arXiv:2606.17386 (2026) 2, 4

    Xiong, Z., Li, W., Wu, Z., Rangesh, A., Bonde, S., Hall, G., Tang, C., Hu, Y., Zhan, W.: TerraTransfer: Learning end-to-end driving policies without expert demonstrations. arXiv preprint arXiv:2606.17386 (2026) 2, 4

    Pith/arXiv arXiv 2026

  35. [35]

    In: NeurIPS (2025) 3

    Yang, J., Chitta, K., Gao, S., Chen, L., Shao, Y., Jia, X., Li, H., Geiger, A., Yue, X., Chen, L.: ReSim: Reliable world simulation for autonomous driving. In: NeurIPS (2025) 3

    2025

  36. [36]

    In: ECCV (2024) 3

    Zhang, C., Biswas, S., Wong, K., Fallah, K., Zhang, L., Chen, D., Casas, S., Urtasun, R.: Learning to drive via asymmetric self-play. In: ECCV (2024) 3

    2024

  37. [37]

    In: ICCV (2021) 4

    Zhang, Z., Liniger, A., Dai, D., Yu, F., Van Gool, L.: End-to-end urban driving by imitating a reinforcement learning coach. In: ICCV (2021) 4

    2021

  38. [38]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2026) 4, 8 18 Y

    Zhou, H., Lin, L., Wang, J., Lu, Y., Bai, D., Liu, B., Wang, Y., Geiger, A., Liao, Y.: HUGSIM: A real-time, photo-realistic and closed-loop simulator for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence (2026) 4, 8 18 Y. Yin et al. A Simulator and Training Details This section details the per-agent reward and conditioning...

    2026