---
license: mit
tags:
  - egoengine
  - dexterous-manipulation
  - egocentric-video
  - flow-matching
  - isaaclab
  - reproduction
---

# egoengine-repro-artifacts

Large artifacts from the independent, from-scratch reproduction of
**EgoEngine: From Egocentric Human Videos to High-Fidelity Dexterous Robot
Demonstrations** (arXiv:2606.12604; no official code exists).

Code + docs + small result artifacts:
https://github.com/physercoe/egoengine-repro

Folders:

- `policy-checkpoints/` — final Appendix-D flow-matching policy checkpoints
  (`best.pt`, `last.pt`, TorchScript export `best_ts.pt`) + normalization
  stats (`stats.npz`, `dataset_meta.json`). ~235 MB.
- `retarget-outputs/` — mink-IK-retargeted bimanual motions for all 16 TACO
  demonstration sequences (`trajectory_kinematic.npz` per sequence +
  `retarget_report.json` + error/QA plots). ~201 MB.
- `rl-checkpoints/` — residual-PPO checkpoints, tensorboard logs and
  `eval_residual.json` results for the escalated refine chunks (local +
  remote A800 batch runs). ~51 MB.

Each folder has its own README with provenance (scripts, configs) and
license notes. See the GitHub repo README for headline results and honest
limitations (sim-only, thin-data policy proof, RL stage under budget).