Reproduce “TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion” on OpenResearch

About this paper

Merkourios Simos, Chengkun Li, Bianca Ziliotto, Alexander Mathis

TERRA infers task-relevant static support geometry from motion kinematics, then retargets the motion to a muscle-actuated model using terrain-interaction and anatomical constraints. A policy trained on 9.4 hours of paired data imitates held-out non-flat locomotion in simulation. It reaches 87.59 ± 0.18% completion, versus 82.16 ± 0.41% for TERRA+Voronoi, but has higher MPJPE in tested simulation, not complete scenes or physical deployment.

Key insights

• Motion trajectories provide both positive evidence of support—such as persistent foot contacts—and negative evidence where the body must be able to move freely; combining these cues improves terrain reconstruction.
• Retargeting constraints materially affect reference quality: the full formulation reduced terrain penetration from 33.45% to 0.16% and raised contact-timing F1 from 88.49% to 93.80%.
• Higher completion and lower tracking error do not always coincide: TERRA had the best overall completion rate, while the Voronoi-terrain variant achieved lower overall joint-position error and performed better on some stair and seat measures.

Summary by alphaXiv.