Robot-level tests of the Robotic World Model: simulators and policies
This repository holds the learned simulators and walking policies from an independent study of the Robotic World Model (RWM; Chenhao Li, Andreas Krause and Marco Hutter, arXiv:2501.10100) and its uncertainty-aware follow-up (RWM-U, arXiv:2504.16680). The study is not affiliated with, endorsed by or reviewed by the original authors, the Robotic Systems Lab or ETH Zurich.
The question. A learned simulator (a "world model") predicts how a robot's state changes from one moment to the next. It can be scored without a robot, for example by how far its predictions drift from recorded data. Do those scores tell us whether a walking policy trained entirely inside the simulator will actually walk?
The answer so far: not settled. Across 19 simulators (the authors' released model and 18 versions of it
fine-tuned by different amounts), the pre-registered tests did not settle. Every fine-tuned version predicted one
step and eight steps ahead more accurately than the released model, yet those two scores ranked walking in reverse
(Spearman's rank correlation about −0.4), and 13 of the 18 versions counted as failed under the registered rule.
Running the authors' own walking policy inside each simulator ranked them well (rank correlation +0.79), but that
was a secondary test. Full results: results/d_stats.txt and README.md in the code repository.
Where the policies were tested. Not on hardware, and not in the authors' Isaac Lab simulator: in a MuJoCo model of the ANYmal D quadruped built from ANYbotics' published description (BSD 3-Clause licence), with a motor model fitted to recorded joint torques.
Related repositories
| What | Where |
|---|---|
| Code, pre-registrations, results and paper for this study | https://github.com/joyjeet-singh/rwm-robot-level |
| The reproduction of RWM's dynamics model that this study builds on | https://github.com/joyjeet-singh/rwm |
| Checkpoints from that reproduction | https://huggingface.co/Joyjeetsingh/rwm-reproduction |
| The authors' released code and model | https://github.com/leggedrobotics/robotic_world_model_lite |
What is here
Every file sits at the path the study's scripts expect (runs/...), so downloading into a clone of the code
repository puts each file where the scripts look for it. 490 files, 596 MB in total.
| Folder | Files | Size (MB) |
|---|---|---|
runs/c/ (simulators, policies) |
70 | 36.9 |
runs/c/d3_lite/ |
6 | 1.2 |
runs/c/lite/ |
15 | 7.1 |
runs/d/ (simulators, run records) |
34 | 214.3 |
runs/d/eval/ |
60 | 4.5 |
runs/d/policies/ |
41 | 6.2 |
runs/ft/ |
30 | 162.6 |
runs/lite/ |
132 | 35.3 |
runs/s1_baselines/ |
36 | 123.4 |
runs/s2/ |
66 | 4.6 |
runs/d/<run>_d<steps>/final.pt: the ladder (Option D). The authors' released model, fine-tuned on data from the MuJoCo robot and saved after 5, 20, 40, 80, 160 or 320 steps (one step is one update on 256 training windows). Run names:Wtrained on walking data (549,773 transitions),Mon mixed data (999,973 transitions, many of them from weak or noisy policies),tftrained one step ahead instead of on 8-step rollouts,ga learning rate ten times smaller (1e-5 instead of 1e-4),s0/s1the seed.runs/d/<run>/run.jsonrecords each run. The 18 saves that were tested are listed inresults/d_stats.txtof the code repository.runs/d/policies/<simulator>__p<k>/: the walking policies (actor.pt) trained inside each tested save, two or three per save, with their training records.runs/d/eval/holds their measurements in MuJoCo.runs/s2/: Step 2. Policies trained inside the released model (REL) and its eight compressed copies, two per simulator:fp16(every number rounded to 16-bit precision),int8w(weights rounded to 8 bits),u10tou50(10 to 50% of the weights set to zero, smallest first),s192ands128(the recurrent layers cut to 192 or 128 units).__p0marks the pilot policies, including those of the rebuilt single-network simulatorsA_s0andB_s0.runs/c/: Option C. Simulators trained from scratch on 10,000 or 6 million MuJoCo transitions, and the policies trained in them (actor_i<iterations>.ptsaved along the way). These simulators failed the walk test.runs/ft/: Study FT. Copies of the reproduction's models with 30% or 50% of their weights set to zero, then fine-tuned for 500 updates with those weights held at zero.runs/s1_baselines/: Study AR. Twelve baseline simulators (MLP, RSSM and Transformer; trained on their own 8-step rollouts,ar, or one step ahead,tf; two seeds each), saved after 500 and 2,500 updates.runs/lite/,runs/c/lite/,runs/c/d3_lite/: the policy trainer's records (summary.json,metrics.jsonl).
Not uploaded: test and smoke runs, resumable training state, diagnostic arrays and the policy trainer's intermediate
checkpoints (every policy's final actor is here as actor.pt). The authors' released model and data are not
redistributed; the code repository fetches them and checks their SHA-256 sums.
Download
from huggingface_hub import snapshot_download
snapshot_download("Joyjeetsingh/rwm-robot-level", local_dir=".", allow_patterns="runs/*")
Run this inside a clone of https://github.com/joyjeet-singh/rwm-robot-level. The .pt files are PyTorch pickle
files: check each against MANIFEST.json (size and SHA-256) before loading, as with any pickle file.
Citation
If you use these files, please cite the original work:
@article{li2025rwm,
title = {Robotic World Model: A Neural Network Simulator for Robust Policy Optimization in Robotics},
author = {Li, Chenhao and Krause, Andreas and Hutter, Marco},
journal = {arXiv preprint arXiv:2501.10100},
year = {2025}
}
@article{li2025rwmu,
title = {Uncertainty-Aware Robotic World Model Makes Offline Model-Based Reinforcement Learning Work on Real Robots},
author = {Li, Chenhao and Krause, Andreas and Hutter, Marco},
journal = {arXiv preprint arXiv:2504.16680},
year = {2025}
}
Licence
Apache License 2.0; see LICENSE. NOTICE lists what these files derive from and what was changed.
Uploaded 17 November 2026 by Joyjeet Singh (independent researcher).