Robot manipulation demonstration datasets + offline learning-from-demonstration framework — a candidate testbed for a continuous-action-space extension of QVIRL/ValueWalk, since it’s real/sim robot manipulation data with continuous control actions, not tabular/gridworld.
Reference
- Title: robomimic — A Framework for Robot Learning from Demonstration
- Team: Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Matthew Bronars, Vaibhav Saxena et al. (Stanford / UT Austin, ARISE Initiative)
- Site: https://robomimic.github.io/
- Docs: https://robomimic.github.io/docs/introduction/overview.html
- Code: https://github.com/ARISE-Initiative/robomimic
- Study paper: “What Matters in Learning from Offline Human Demonstrations for Robot Manipulation” — https://arxiv.org/abs/2108.03298 (arXiv 2108.03298, 2021)
In one line
robomimic bundles a broad set of demonstration datasets (human teleop + machine-generated, across several robot manipulation tasks) together with offline learning algorithms, aimed at making offline learning-from-demonstration research on continuous-control manipulation more accessible.
My notes
TBD — fill in after digging into the docs/datasets (task list, action space dims, human vs. machine demos, MuJoCo/robosuite sim backend).
Relevance to thesis
Potential source of real continuous-action demonstration data for validating the continuous-action-space extension of QVIRL/ValueWalk beyond synthetic/tabular benchmarks — offline, demonstration-based, exactly the regime BIRL methods like AVRIL target. Would need to check whether reward/value supervision (vs. just state-action trajectories) is available or would need to be inferred, and how the action space dimensionality compares to what NAF / DDPG-style continuous control assumes.
Questions for Ondřej
- Do the robomimic tasks have known/ground-truth rewards to validate a learned reward posterior against, or would evaluation have to be purely behavioral (policy performance)?