Robotics · 2026-11-03
Right Ranking, Wrong Scale: Reproducing the Robotic World Model's Dynamics Model and Testing the Uncertainty of Its Follow-up
Joyjeet Singh
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors reimplement the Robotic World Model (RWM) dynamics model and test RWM-U’s ensemble-based uncertainty used to penalize reward in offline policy optimization. Their rebuild matches the released model’s gradients. On held-out data, RWM’s rollout training beats teacher forcing at a preregistered long horizon for all ten episodes, but loses at one step. RWM-U ranks rollout errors correctly (Spearman ρ=0.90 at 100 steps) yet its spread is badly miscalibrated and too small; they also find defects in the released evaluation pipeline.
Why it matters
This work isolates whether RWM-U’s ensemble disagreement is a trustworthy “trust metric” by testing both ranking and scale via held-out uncertainty calibration/coverage. It shows ensemble disagreement can correlate with error while still severely underestimating uncertainty, and it identifies specific pipeline defects. It also provides a reproduced dynamics-model implementation and gradient-match verification, enabling more reliable follow-up research on world models and offline model-based control.
Method
- Reimplemented RWM’s recurrent (GRU) ensemble dynamics model from paper/released descriptions; verify wiring by matching released gradients before training (bitwise inference outputs; max gradient difference 0.000e+00).
- Evaluate rollout-training vs teacher forcing on a held-out protocol with each of ten episodes held out once; test long-horizon and one-step windows using the authors’ metrics and bootstrap/sign tests.
- Test RWM-U ensemble uncertainty as ranking vs calibration: compute Spearman rank correlation, error-to-spread ratio, and ±kσ coverage at multiple horizons; study causes of under-coverage and test per-horizon spread rescaling and pipeline defects.
Limitation
The study covers one robot, one gait and one dataset; it trains no policy.
Abstract (from arXiv)
We reimplement the dynamics model of the Robotic World Model (RWM) and test the uncertainty estimate that its follow-up, RWM-U, uses to penalize reward in offline policy optimization. Our implementation matches the released model's gradients. On held-out data, training on the model's own rollouts beats one-step training at the preregistered long horizon for all ten episodes, but loses on short one-step windows. RWM-U's ensemble disagreement ranks rollout errors better than forecast step (Spearman rho = 0.90 at 100 steps), yet its spread substantially understates error: one-sigma coverage on the released checkpoint is 4.6% at 100 steps, versus 68% for a calibrated Gaussian. Rolling ensemble members on their own predictions raises coverage to 33%. The implemented squared-error loss on sampled predictions drives each member's variance toward zero. A horizon-specific rescaling restores one-sigma coverage under leave-one-episode-out evaluation, but not two-sigma coverage. We also identify two defects in the released evaluation pipeline. The study covers one robot, one gait and one dataset; it trains no policy.
Related papers
- Rollback: Experience Backtracking in World Models for Policy Improvement
- AFTER: A World Model with Adaptive Fast--Slow Timescales for Error-Driven Retention
- Causal world models under physical parameter shifts
- How Many Candidates Does a One-Step World Model Need? A Random-Codebook Account of Nearest-of-m Training on Frozen DINO Tokens
- LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-03 · Indexed by Pub-AI: 2026-11-03 UTC
Right Ranking, Wrong Scale: Reproducing the Robotic World Model's Dynamics Model and Testing the Uncertainty of Its Follow-up
First announced by arXiv: 2026-11-03 · Original record ↗
- arxiv
- 2611.02699
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-03 · First detected: 2026-11-03 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.02699v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
We reimplement the dynamics model of the Robotic World Model (RWM) and test the uncertainty estimate that its follow-up, RWM-U, uses to penalize reward in offline policy optimization. Our implementation matches the released model's gradients. On held-out data, training on the model's own rollouts beats one-step training at the preregistered long horizon for all ten episodes, but loses on short one-step windows. RWM-U's ensemble disagreement ranks rollout errors better than forecast step (Spearman rho = 0.90 at 100 steps), yet its spread substantially understates error: one-sigma coverage on the released checkpoint is 4.6% at 100 steps, versus 68% for a calibrated Gaussian. Rolling ensemble members on their own predictions raises coverage to 33%. The implemented squared-error loss on sampled predictions drives each member's variance toward zero. A horizon-specific rescaling restores one-sigma coverage under leave-one-episode-out evaluation, but not two-sigma coverage. We also identify two defects in the released evaluation pipeline. The study covers one robot, one gait and one dataset; it trains no policy.
