TERRA Brings Muscle Control Beyond Flat Ground
EPFL’s TERRA reconstructs terrain from motion capture, retargets movement under anatomical and contact constraints, and trains muscle-actuated policies for stairs, ramps, platforms, seats, and uneven terrain. The research provides a practical path from scene-less motion data to terrain-aware embodied control ([paper](https://arxiv.org/abs/2609.38653), [project](https://cnai.epfl.ch/terra/)).
TERRA’s key contribution is turning missing environment labels into usable supervision for physically grounded locomotion.
- –Reconstructs support geometry from kinematics, contact estimates, terrain priors, and negative free-space evidence
- –Combines anatomical, tendon-continuity, contact, clearance, and collision constraints during retargeting
- –Trains one policy across diverse terrain families using 9.4 hours of motion data from five datasets
- –Moves musculoskeletal simulation beyond flat-ground imitation toward stairs, ramps, beams, stepping stones, and uneven terrain
- –The main caveat is sim-to-real transfer: the results remain bounded by simulation assumptions and supported terrain categories
DISCOVERED
1h ago
2026-10-11
PUBLISHED
1h ago
2026-10-11
RELEVANCE
AUTHOR
AI Search