IRL-VLA: Training an Vision-Language-Action Policy via Reward World Model
1. Paper overview
In one sentence: IRL-VLA replaces simulator reward queries during policy refinement with a learned trajectory scorer, yielding a small aggregate planning gain while shifting the balance between progress, safety and comfort. motivationpolicyreward-datareward-modelrlmain-resultschallenge
| At a glance | What to know |
|---|---|
| Research problem | Source description Imitation restricts driving policies to recorded behavior, while sensor-rendering simulators make exploration expensive. The paper seeks multi-modal planning under competing safety, progress and comfort objectives using learned reward feedback. motivation |
| Core mechanism | Source description A three-stage pipeline couples a semantic/geometric diffusion VLA to a separately trained Reward World Model (RWM). policyarchitecturereward-model |
| A key reported result | Navhard trajectory planning: IRL-VLA-RL 74.9; IRL-VLA-PT 74.4. EPDMS (higher is better). NAVSIM non-reactive evaluation; stated 1,192 navtrain/136 navhard scenes; Table 1 labels all entries Stage 1. GTRS-Aug 74.3; WOTE 66.7; DiffusionDrive 63.2. Input resolutions and backbones differ. RL adds 0.5 score points over PT. EP rises 83.9→96.2, while NC falls 98.3→96.9, DAC 92.4→91.3 and EC 76.0→72.4. This is a tradeoff, with no reported uncertainty. main-resultsevaluation |
| Reading caution | Reader analysis The paper supplies neither direct RWM accuracy/calibration results nor a controlled RWM-versus-simulator RL comparison. Its tables do not quantify speedup, domain-gap reduction or robustness to reward exploitation. reward-modelmain-resultscomponentsweight-ablation |
Core contributions
- Source description
A three-stage pipeline couples a semantic/geometric diffusion VLA to a separately trained Reward World Model (RWM). policyarchitecturereward-model
- Reader analysis
The authors describe simulator-free, closed-loop VLA reinforcement learning. This applies most clearly to policy-update feedback: simulator-generated labels remain necessary for RWM training. motivationreward-datarl
Figure 1. The proposed shortcut changes where policy-training feedback comes from. Original paper, p. 1 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the rows from top to bottom, following the arrows entering the green policy box. In row (a), demonstrations supply supervision. In row (b), the policy sends actions to a simulated environment and receives rewards; simulated sensor observations also feed the policy. In row (c), real sensor data feed both the policy and the red Reward World Model. The bottom arrow carries the policy's proposed action into that model, while the upper arrow returns reward feedback. The purple demonstrations box represents reward learning at a conceptual level. Section 3.3 adds an essential detail absent from this row: simulator-computed metric labels supervise the reward model. motivationreward-datareward-modelarchitecture
What it supports. The proposal reduces dependence on expensive simulator queries inside the policy-update loop by learning their reward consequences. It retains a distinct policy and reward predictor, which supports the catalog's Dual-system classification. The graphic's light-computation and online-exploration badges are author claims, not measurements of speed or deployment performance.
Where the evidence stops. The demonstration-to-RWM arrow is incomplete as a data-provenance account: Section 3.3 explicitly uses simulator feedback. Consequently, the diagram does not support a claim that the entire training pipeline is simulator-free.
2. Motivation
2.1 The problem and the proposed response
Imitation restricts driving policies to recorded behavior, while sensor-rendering simulators make exploration expensive. The paper seeks multi-modal planning under competing safety, progress and comfort objectives using learned reward feedback. motivation
2.2 What this reading follows
A driving policy needs feedback about alternatives that human demonstrations never selected. IRL-VLA supplies that feedback through a separate Reward World Model: first imitate driving trajectories, then learn to predict simulator-assigned scores, and finally improve the diffusion policy using those predictions. The useful distinction is between learning a trajectory's score and simulating the future world. This paper specifies the former much more clearly. Its figures explain the training loop, while its tables show how the resulting policy redistributes driving metrics. Read the 74.9 Navhard result alongside its 74.4 imitation baseline and the component scores, keeping the separately reported challenge result apart. motivationpolicyreward-datareward-modelrlmain-resultschallenge
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
Dual-system / Other mechanisms / Outside quadrants fits a separate trajectory policy and reward predictor coupled through RL feedback. The reward model predicts metric consequences; there is no demonstrated joint future-observation/action generator or inverse-dynamics action extraction. The policy's internal semantic/geometric branches alone are not the basis for this judgment. policyarchitecturereward-modelrl
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 1. Build a policy that can propose more than one plausible plan
Begin with the distinction between scene meaning and scene geometry. The Senna-VLM-based branch supplies semantic information from images and language, while BEV features and detection/map tokens provide geometric context. The diffusion planner uses both, plus the driving command, to denoise anchored trajectory proposals. Its output is a sequence of ego-frame position and heading waypoints. During imitation, reconstruction and anchor classification losses connect these proposals to demonstrations. Reader interpretation: this creates a useful starting distribution for later reward optimization, because reward feedback cannot improve alternatives that the policy never proposes. Table 2 tests the initial architecture by adding semantics and diffusion sequentially. It supports those additions under the reported setup, without isolating every possible component interaction or demonstrating low-level vehicle execution. policyimitationarchitecturecomponents
Figure 2. Scene-conditioned trajectory generation and trajectory scoring remain separate modules. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the stage letters rather than the physical left-to-right order: (a) imitation, (b) reward learning at the right, then (c) reinforcement learning in the center. In (a), semantic reasoning, 3D reasoning and command encoding condition the unified diffusion policy. In (b), a trajectory encoder adds the candidate plan to the scene information, and simulator feedback supplies the reward target. The center loop sends trajectories downward to the reward model and returns rewards upward to the policy. The lower panels reveal different outputs: denoised trajectories in (f), trajectory rewards in (g). Section 3.2 confirms the displayed labels (d) semantics, (e) 3D reasoning and (f) diffusion. architecturepolicyreward-datareward-modelrl
What it supports. The RWM learns to judge proposed trajectories in their scene context; the policy learns to generate them. PPO connects the two during refinement. The drawing therefore supports reward-guided policy post-training, but does not establish a shared model that jointly generates future observations and executable actions.
Where the evidence stops. The caption incorrectly assigns diffusion and semantics to panels (e)/(f); the diagram and Section 3.2 agree on (f)/(d). Algorithm 1 also invokes an unspecified next-state model, so the feedback arrows do not establish a complete reactive simulator.
5.2 2. Learn a consequence score without generating the next camera frame
The RWM receives a scene and a candidate trajectory, extracts BEV features at its waypoints, and predicts eight metric components. Its training set broadens beyond final imitation outputs by retaining intermediate diffusion trajectories, varying K-means anchor sets and changing simulated ego poses. The simulator supplies the target scores. Reader interpretation: the documented learning mechanism is supervised metric prediction, even though the paper calls it inverse reinforcement learning. This distinction matters because the model approximates a predefined evaluator rather than establishing that it recovered an unknown expert preference. There is also an objective mismatch to preserve: the predicted training reward is written as a weighted sum excluding EC, whereas evaluation Equation (7) combines multiplicative penalties with a weighted average including EC. A reproduction must decide how the implemented scorer reconciles these descriptions. reward-datareward-modelimplementationevaluation
5.3 3. Improve the learned reward and then check what actually improved
After pretraining, sampled diffusion trajectories are evaluated by the RWM. The method combines reward and critic information through GAE, standardizes advantages within groups, and uses them to update the policy while retaining a weighted imitation loss. The source names PPO, but its explicit loss and generic algorithm are not fully reconciled, so implementation fidelity requires clarification. Reader interpretation: optimizing a learned score creates two distinct questions—whether the policy raises that score and whether simulator-measured driving improves. Table 1 answers only the latter at the reported endpoints: aggregate EPDMS increases slightly while progress rises and several safety/comfort scores fall. Table 3 motivates retaining imitation, but does not explain its claimed low-weight collapse. Neither table measures reward-model calibration as the policy's trajectory distribution changes. rlmain-resultsweight-ablationreward-model
5.4 Training and inference
During training
Imitation pretraining uses V2-99, 256×704 images, 100 epochs, AdamW at learning rate 10^-4 and batch size 32. Reported PPO settings are clipping 0.2, discount 0.99 and GAE parameter 0.95; training uses eight NVIDIA A100 GPUs. implementation
Section 4 specifies BCE for binary reward heads, MSE for continuous EP and cross-entropy for ternary NC/DDC. Section 3.3 instead displays a weighted error-norm loss; their implementation relationship is unresolved. reward-modelimplementation
During inference
The optimized policy maps scene/ego/command features to waypoint trajectories through conditional denoising. RWM supplies training feedback; an inference-time RWM reranker or executable low-level controller is not specified. policyrl
Algorithm 1 invokes a predicted next-state transition without defining its implementation. The following formulation treats denoising as an intrinsic MDP. This does not establish a learned renderer or fully reactive driving environment. rlevaluation
5.5 Implementation flow
- Encode semantics and geometry
A Senna-VLM-based branch processes images and language. A BEV encoder/adaptor supplies scene features, detection tokens and map tokens. Command features also condition planning. policyarchitecture
- Denoise anchored trajectories
Gaussian-noised anchor proposals attend to semantic, BEV, map and object information. An MLP decodes refined trajectory tokens; imitation learns L1 reconstruction and binary cross-entropy anchor classification. policyimitation
- Collect reward supervision
Score intermediate denoising trajectories with NAVSIM. Diversify candidates through K-means anchor sets, K=32–8192, and multiple simulated ego poses. Extended Comfort (EC) is omitted from RWM supervision because it requires two simulations per scene. reward-data
- Learn trajectory rewards
Sample BEV features along candidate waypoints and predict eight metric components with separate heads. Reader analysis: despite the inverse-RL label, the specified objective directly fits simulator-provided targets; recovery of an unknown expert reward is not demonstrated. reward-modelimplementation
- Refine the policy
Evaluate sampled diffusion trajectories with the RWM, combine rewards with critic estimates, compute generalized advantage estimation (GAE) and group-standardized advantages, then update the policy using the stated Proximal Policy Optimization (PPO) procedure while retaining imitation loss. rl
6. Experiments & results
IRL-VLA trains an autonomous-driving diffusion policy with a separate model of trajectory rewards. Imitation initializes the policy, simulator-labeled trajectories train the reward model, and reinforcement learning refines the policy. The reported Navhard gain is small in aggregate and redistributes progress, safety and comfort scores. The evidence concerns simulated planning rather than physical driving.
6.1 Read the original evidence
Table 1. The aggregate improvement accompanies a substantial change in component scores. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the final two rows to compare imitation pretraining (PT) and reward-guided refinement (RL). Both list V2-99 and 256×704 input resolution, so this contrast is more directly relevant to post-training than comparisons across different backbones. Read the EPDMS column, then move left to EP (ego progress), NC (no at-fault collision), DAC (drivable-area compliance) and EC (extended comfort). These columns reveal which behaviors move with the aggregate. Equation (7) on page 7 describes the evaluation aggregation. The Stage column says Stage 1 even for the RL row; it should not be used to rename the paper's three learning stages. main-resultsevaluationchallenge
What it supports. EPDMS increases from 74.4 to 74.9, a gain of 0.5 score points. EP rises from 83.9 to 96.2, while NC decreases from 98.3 to 96.9, DAC from 92.4 to 91.3 and EC from 76.0 to 72.4. The table supports a better aggregate score with competing component changes.
Where the evidence stops. Baseline resolutions and backbones differ, and no uncertainty is reported. NAVSIM uses non-reactive simulation. The separately stated challenge score of 45.0 has a different reporting context and cannot be substituted into this table.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Navhard trajectory planning NAVSIM non-reactive evaluation; stated 1,192 navtrain/136 navhard scenes; Table 1 labels all entries Stage 1. | IRL-VLA-RL 74.9; IRL-VLA-PT 74.4. EPDMS (higher is better) | GTRS-Aug 74.3; WOTE 66.7; DiffusionDrive 63.2. Input resolutions and backbones differ. RL adds 0.5 score points over PT. EP rises 83.9→96.2, while NC falls 98.3→96.9, DAC 92.4→91.3 and EC 76.0→72.4. This is a tradeoff, with no reported uncertainty. main-resultsevaluation |
| Hierarchical VLA component ablation NAVSIM navhard-real; imitation-trained architecture variants in Table 2. | 3D only 70.0; plus semantics 71.4; plus diffusion 74.4. EPDMS | Sequential gains are 1.4 and 3.0 score points. Supports the combined architecture under this setup; the study is not factorial. Full-model LK differs between Tables 1 and 2 (98.9 versus 99.6). components |
| Imitation weight during reinforcement learning Table 3 and Section 5.3 Navhard ablation; w_IL varied. | w_IL=1.0: 73.9; 0.5: 74.9; 0.1: 73.4. EPDMS | 0.5 leads the three reported settings. The prose calls 0.1 training collapse, but a finite competitive score remains in the table; collapse severity cannot be inferred. weight-ablation |
| CVPR 2025 NAVSIM v2 challenge Author-reported challenge entry, distinct from Table 1 Navhard evaluation. | 45.0; first runner-up. Challenge EPDMS, labeled inconsistently in the prose | No challenge competitor table or detailed protocol is supplied. Retain as an author-reported result; do not merge its score with the 74.9 Navhard result. challenge |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. Sequential additions improve the imitation-trained policy before RWM-guided RL. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the checkmarks and crosses to identify the comparison before reading the shaded EPDMS column. Row 1 includes 3D reasoning alone; row 2 adds semantic reasoning; row 3 also includes the diffusion planner. Section 5.3 describes these as models trained from human driving demonstrations, so the table addresses the strength of the initial VLA architecture. It is not a reward-model accuracy test. The incremental scores should be interpreted in this order: the effect of semantics is measured without diffusion, while the effect of diffusion is measured with semantics already present. Other combinations are not shown. componentspolicymain-results
What it supports. Semantic reasoning raises EPDMS from 70.0 to 71.4, and the subsequent diffusion addition raises it to 74.4. This supports the combined policy design on the reported navhard-real setting. It does not isolate every interaction between the components or explain the additional RL gain.
Where the evidence stops. Full-model LK is 99.6 here but 98.9 in the PT row of Table 1. The source gives no reconciliation. Preserve both values and avoid using them as interchangeable measurements of the same configuration.
Table 3. Retaining imitation supervision matters, but the reported collapse remains unexplained. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The first column varies the coefficient w_IL in Equation (6), which adds imitation loss to the reinforcement objective. Read the three EPDMS entries as a small hyperparameter sweep, then compare the component columns if investigating why the scores change. The middle row, w_IL=0.5, matches the reported RL model's aggregate score. A coefficient of 1.0 means an unscaled additive imitation term; it does not by itself establish equal gradient contributions from the two losses. The authors call the 0.1 setting a training collapse in Section 5.3, so read this table together with that prose rather than treating the label as a measured failure criterion. weight-ablationrl
What it supports. Among the three settings, w_IL=0.5 reaches 74.9 EPDMS, compared with 73.9 at 1.0 and 73.4 at 0.1. The table supports sensitivity to imitation regularization and a best reported setting of 0.5. It provides neither a universal optimum nor a quantified collapse threshold.
Where the evidence stops. The 0.1 row still reports 73.4 EPDMS. No training curve, seed distribution or collapse definition explains the stronger prose claim. The PPO/KL formulation ambiguity also limits exact reproduction of this sweep.
7. Analysis & limitations
7.1 What the evidence leaves open
The paper supplies neither direct RWM accuracy/calibration results nor a controlled RWM-versus-simulator RL comparison. Its tables do not quantify speedup, domain-gap reduction or robustness to reward exploitation. reward-modelmain-resultscomponentsweight-ablation
Non-reactive NAVSIM results do not establish physical deployment, reactive traffic interaction or broad generalization. No dedicated limitations section, seeds or confidence intervals accompany the reported results. evaluationmain-resultscomponentsweight-ablation
Figure 2's caption mislabels detailed panels e/f relative to the diagram and Section 3.2. Interpret the labeled modules using the matching method text; do not silently repair the source graphic. architecture
7.2 Questions for discussion
- Does improving learned reward preserve simulator reward on newly explored trajectories?
- How would including EC and matching Equation (7)'s aggregation change the progress/safety tradeoff?
- What implementation reconciles Algorithm 1, Equation (5), and the claimed low-imitation training collapse?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires camera/ego/command processing, Senna-VLM, V2-99 BEV features, anchor generation, NAVSIM metric labels, reward heads and diffusion RL/critic training. The source omits exact checkpoints, frozen-module choices, selected anchor count, denoising schedule, horizon, RWM/RL training budgets and software versions. Numeric metric weights are delegated to a cited work rather than enumerated here. policyreward-datareward-modelrlimplementation
Proposed checks: compare learned versus simulator rewards on held-out candidates before and after RL, and repeat the imitation-weight sweep across seeds while tracking simulator score and objective components. These would test reward exploitation and the claimed collapse. reward-datareward-modelrlweight-ablation
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Check whether RL exploits the reward approximation
Reader-proposed experiment, not performed: hold out complete NAVSIM scenes from reward fitting and collect matched candidate sets from the PT and RL policies. Score every candidate with both the RWM and the simulator. Compare per-metric error, trajectory-ranking agreement and the simulator score of the RWM-selected candidate, reporting EC separately because the RWM excludes it. Keep scenes and candidate counts fixed across policies. If learned reward improves after RL while simulator score or ranking agreement worsens, that falsifies reliable reward transfer on this test and indicates exploitation or distribution shift. reward-datareward-modelevaluationmain-results
Check 2: Make the imitation-weight collapse claim falsifiable
Reader-proposed experiment, not performed: repeat w_IL=1.0, 0.5 and 0.1 from the same PT checkpoint with a fixed RWM, identical update budgets and multiple seeds. Declare the exact PPO objective and KL sign before running. Track learned reward, simulator EPDMS, imitation loss, policy divergence and gradient magnitudes throughout training; define collapse in advance as sustained non-finite optimization or sustained simulator-score degradation. Stable runs near the table's 73.4 at w_IL=0.1 would not support a general collapse claim; repeated degradation would identify when and how it occurs. rlimplementationweight-ablationmain-results
8.3 Reading coverage
Visual audit: Visually inspected the title, all 14 authors, affiliations and v3 margin on p. 1; motivation on p. 2; policy and waypoint formulation on p. 3; Figure 2, its caption, imitation loss, reward-data construction and Table 1 on p. 4; reward equations and Algorithm 1 on p. 5; diffusion/PPO equations, all stated training settings and dataset split on p. 6; evaluation Equation (7), Table 2, ablation prose and challenge claim on p. 7; Table 3 and conclusion continuation on p. 8. All five final crops were viewed at original resolution and retain complete diagrams or table headers/rows without long captions. Arrow directions were checked against the method: trajectories enter the RWM and reward returns to the policy. Figure 2's caption-label discrepancy, reward/evaluation aggregation mismatch and Table 3/prose discrepancy are disclosed. All four text chunks, including references on pp. 9–10, were read. Reference pages were not visually inspected; no appendix is present and separate supplements remain unverified.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title/author block and Abstract (p. 1)
- 1. Introduction (pp. 1–2)
- 2. Related Work (pp. 2–3)
- 3. Method, Sections 3.1–3.4 and Algorithm 1 (pp. 3–6)
- 4. Implementation Details (p. 6)
- 5. Experiments, Sections 5.1–5.3 (pp. 6–7)
- 6. Conclusions and final ablation table (pp. 7–8)
- References (pp. 9–10)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Identity/version note: the supplied main title and all 14 authors match p. 1. The printed title additionally continues 'for End-to-End Autonomous Driving'. The artifact is arXiv:2508.06571v3, dated 15 Aug 2025, later than the catalog submission date of 7 Aug 2025. Earlier editions and revision history were not supplied; no claim of edition equivalence is made.
- Acquisition caveat retained: text extraction does not reconstruct figure images. Both original figures and all three tables were therefore inspected in the PDF and as final crops; supporting pages 1–8 were visually inspected.
- Separate supplemental material availability has not been fully verified. No appendix is present in the supplied ten-page PDF.
- References on pp. 9–10 were read as text; cited works were not acquired. The linked project repository was not inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title/author block and arXiv margin
The main title and all 14 authors match the catalog. The title continues 'for End-to-End Autonomous Driving'; the artifact is arXiv:2508.06571v3, 15 Aug 2025. Five affiliations are printed.
Go to primary source ↓motivationPDF pp. 1–2, Abstract, Figure 1 and Section 1
The authors propose replacing expensive simulator feedback during VLA policy optimization with a learned reward model after imitation pretraining.
Go to primary source ↓policyPDF p. 3, Sections 3.1–3.2; p. 4, Figure 2(a,d,e,f)
Senna-VLM semantics, BEV geometry with map/detection tokens, command encoding and anchored diffusion yield ego-frame position/heading waypoints.
Go to primary source ↓architecturePDF p. 4, Figure 2(a–g) and caption; p. 3, Section 3.2
Stages a/b/c are arranged a/c/b. Diagram d/e/f denote semantics/3D/diffusion, agreeing with Section 3.2; the caption misidentifies e/f.
Go to primary source ↓imitationPDF pp. 3–4, Section 3.2, Imitation Policy Learning Loss
The decoder predicts anchor scores and denoised trajectories using L1 reconstruction and binary cross-entropy classification.
Go to primary source ↓reward-dataPDF p. 4, Section 3.3, Reward Data Collection
Simulator scores label intermediate diffusion trajectories; K-means uses K from 32 to 8192, with multiple ego poses. EC requires two simulations per scene and is excluded.
Go to primary source ↓reward-modelPDF p. 5, Section 3.3, reward heads, aggregation and optimization equations
Waypoint-sampled BEV features feed eight heads. EP is continuous; DAC/TLC/TTC/LK/HC binary; NC/DDC ternary. Predicted rewards are weighted sums; the displayed fitting loss uses weighted error norms.
Go to primary source ↓rlPDF p. 5, Section 3.4 and Algorithm 1; p. 6, Equations (1)–(6)
RWM/critic feedback supports GAE, standardized advantages and diffusion updates with imitation. Algorithm 1 invokes an unspecified transition model; Equation (5) lacks an explicit clipped PPO ratio and prints negative KL.
Go to primary source ↓implementationPDF p. 6, Section 4
V2-99, 256×704 images; imitation uses 100 epochs, AdamW, learning rate 10^-4, batch 32. Reward losses are BCE/MSE/CE by type. PPO clipping is 0.2, discount 0.99, GAE 0.95; eight A100 GPUs.
Go to primary source ↓evaluationPDF pp. 6–7, Section 5.1 and Equation (7)
Stated split: navtrain 1,192 scenes, navhard 136 scenes. NAVSIM is non-reactive. EPDMS multiplies NC/DAC/DDC/TLC penalties by a normalized weighted TTC/EP/HC/LK/EC average.
Go to primary source ↓main-resultsPDF p. 4, Table 1, PT/RL and baseline rows; p. 7, Section 5.2
PT/RL: EPDMS 74.4/74.9, EP 83.9/96.2, NC 98.3/96.9, DAC 92.4/91.3, EC 76.0/72.4. GTRS-Aug/WOTE/DiffusionDrive score 74.3/66.7/63.2 EPDMS; resolutions/backbones vary.
Go to primary source ↓componentsPDF p. 7, Table 2 rows 1–3 and Section 5.3; p. 4, Table 1 PT row
3D only / plus semantics / plus diffusion score 70.0/71.4/74.4 EPDMS. Full-model LK is 99.6 in Table 2 versus 98.9 for PT in Table 1.
Go to primary source ↓weight-ablationPDF p. 8, Table 3 all rows; p. 7, Section 5.3, Imitation Loss Weight
Weights 1.0/0.5/0.1 give 73.9/74.9/73.4 EPDMS. Prose calls 0.1 training collapse without a curve or criterion reconciling it with the table.
Go to primary source ↓challengePDF p. 2, Section 1; p. 7, Section 6
Authors report 45.0 and first runner-up in the CVPR 2025 NAVSIM v2 challenge, separately from Navhard 74.9; no challenge breakdown is supplied.
Go to primary source ↓8.5 Primary sources
IRL-VLA: Training an Vision-Language-Action Policy via Reward World Model ↗
PDF · 5,998 extracted words
Source fingerprint
3320349333add6a78611dc4f02230748b9e903684667afccc102e79a8abc13ef