CoRLSample
Cross-Embodiment Transfer in General Manipulation Policies
S. Varga, L. Mbeki, T. Ishikawa — Google DeepMind
Abstract
Physical interaction data cannot be scraped, and collecting it per platform bounds progress in robot learning. We train a manipulation policy on a single platform and evaluate zero-shot transfer to platforms differing in degrees of freedom, end-effector geometry and sensor configuration. Degradation is smaller than prior work predicts. We analyse which representational choices carry the transfer and which do not.
Key findings
- Zero-shot transfer with limited degradation across differing morphologies
- Transfer carried primarily by object-centric rather than joint-space representations
- Reduces the per-platform data requirement by a large factor
- Degradation concentrates in tasks requiring precise force control
Limitations
Published at equal prominence to the findings. A paper’s limitations are usually the part that determines how much its result should move your beliefs.
- All evaluation conducted in laboratory conditions
- Field reliability not measured and not claimed
- Task set excludes deformable objects