- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
Paper introduces federated learning for WAMs without sharing demos or weights. Clients upload DINOv3 features, a server fits a teacher to them, and each client distills the teacher back into its own world model and retrains its action expert. Claims 2 pp in sim, mostly on unseen domains, and +5 pp on every real robot task.
Critical strengths
- The design makes sense, sharing chunk features instead of weights means you don't need matching architectures, and the self-distillation baseline (fit a teacher only from your own predictions) is the right control for whether the gain comes from other clients or just from the distillation step.
- The authors are upfront about what's missing.
Critical weaknesses
- The paper says its own results might not be real. Section 5 calls them "author-supplied point estimates" with a "source-provenance conflict". Logs, configs, checkpoints, trial counts, seeds are all listed as outstanding in appendix. That covers every table and figure in the paper. The only experiments that do have records (Appendix K) don't test the actual method.
- The robot numbers look made up. In Table 5, every Pen-to-Basket cell is exactly Corn-to-Plate minus 5.0, and every Cube-to-Bowl cell is exactly minus 10.0. Fig 15's gain matrices are identical across all three tasks. Real robot evals usually differ significantly across different pick-and-place tasks.
- no seeds or error bars anywhere in the main study
Questions for authors
- Can you clarify the provenance situation in Appendix H? Specifically, which of the reported results come from completed runs of the method in Section 4?
- For the Piper experiments, how many trials per cell were run, and is there an explanation for the constant 5 pp offset between the three tasks in Table 5?
- How many seeds were used for the sim experiments, and could you report per-seed results?