Source note
dWorldEval: Scalable Robotic Policy Evaluation via Discrete Diffusion World Model
Summary
dWorldEval is a robot policy evaluator built as a discrete diffusion world model. It predicts future observations and a task-progress token from language, images, and actions in one token sequence, with the goal of ranking robot policies without running them in every environment.
Problem
- Evaluating robot policies at large scale is expensive with real rollouts and asset-heavy simulators, especially across many tasks and environments.
- Prior world-model evaluators often fail as proxies because they do not follow robot actions well under out-of-distribution or failed behaviors, and they drift over long horizons.
- This matters because policy ranking becomes unreliable when the model hallucinates success, misses failures, or breaks spatial and temporal consistency.
Approach
- The model maps vision, language, and robot action chunks into one discrete token space and predicts them with a single transformer denoiser, so actions are first-class inputs rather than weak side conditions.
- It uses masked discrete diffusion to reconstruct future tokens, jointly generating the next visual state and a discrete progress token.
- A sparse keyframe memory stores low-resolution history frames with explicit frame-index tokens to reduce long-horizon drift and keep object layout consistent.
- During training, task progress is converted into discrete text-like tokens using milestone-based scoring; at inference, success is declared when the predicted progress reaches 1.
- The paper also introduces Δ-LPIPS, a metric that scores how well predicted state changes match ground-truth state changes under the given action sequence.
Results
- On LIBERO action controllability, dWorldEval reports the best Δ-LPIPS on both expert and failure subsets: 0.315 / 0.352, compared with WorldEval 0.423 / 0.701, WorldGym 0.347 / 0.650, and Ctrl-World 0.334 / 0.416. Standard LPIPS is also slightly better: 0.215 vs 0.262 / 0.218 / 0.220.
- For long-horizon round-trip consistency, the full model beats its no-memory ablation across action horizons H=5,10,15,20 with LPIPS 0.130, 0.145, 0.193, 0.243 versus 0.177, 0.186, 0.302, 0.411.
- In the full consistency comparison against prior methods, dWorldEval again has the lowest round-trip LPIPS: at H=20, 0.243 vs WorldEval 0.531, WorldGym 0.482, Ctrl-World 0.370.
- As a policy-evaluation proxy, estimated success rates correlate strongly with real execution: r=0.910 on LIBERO multi-view, r=0.927 on RoboTwin, and r=0.918 on real-world tasks. The introduction also states Pearson r ≈ 0.9 overall.
- On LIBERO single-view ranking, the paper reports minimal rank violation for dWorldEval with MMRV = 0.013, while baselines reach up to 0.039.
- The experiments use LIBERO (5.5k expert demos plus 1k failed rollouts), RoboTwin (5.5k trajectories across 10 tasks), and a real bimanual robot dataset with 5.2k trajectories including 1k failures across 5 tasks.