ICWM + Reflective VLA (code)
Source code for two LeRobot-based policy implementations used in the LIBERO tvshift / camera-shift experiments. Code only: no checkpoints, training outputs, virtualenvs or git history.
| file | contents |
|---|---|
icwm-code.zip |
ICWM, In-Context World Modeling for Robotic Control (arXiv 2606.26025): lerobot/ fork with the icwm policy (Qwen2.5-VL backbone, FAST action tokens, context-clip sampler, cache builder), training/resume scripts in scripts/, evaluation summaries in evaluate_results/ |
reflective-va-code.zip |
Reflective VLA, In-Context Action Consequences (arXiv 2606.25215): plugin package src/reflective_vla on top of LeRobot pi0.5 (block-causal mask, observation-action-consequence triplets), third_party/lerobot, configs, scripts, tests, PLAN.md, evaluation summaries |
See README.md inside each zip for setup and usage.
Related: dinyu/tvshift-wam-idm-wan22, dinyu/tvshift-libero-lerobot, dinyu/libero-raw-hdf5