TERRA: Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion
Merkourios Simos, Chengkun Li, Bianca Ziliotto, Alexander Mathis
EPFL
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 33 · arXiv 2609.38653
The abstract
Recent advances in musculoskeletal modeling and reinforcement learning have enabled muscle-actuated agents to reproduce increasingly complex human motions. Yet these capabilities remain largely confined to flat ground, in part because motion datasets rarely include aligned terrain geometry and because retargeting terrain interactions to complex musculoskeletal bodies is challenging. We present TERRA, an end-to-end pipeline for terrain-aware retargeting and control of musculoskeletal locomotion. From kinematic trajectories alone, TERRA combines terrain priors, estimated contacts, and negative free-space evidence to recover task-relevant support geometry. TERRA further considers anatomical, tendon-continuity, and contact constraints during retargeting. Using the resulting motion-terrain pairs from five datasets, we successfully train a single muscle-actuated control policy on 9.4 hours of diverse locomotion. Across reconstruction, retargeting, and held-out tracking benchmarks, TERRA improves terrain accuracy, sharply reduces anatomical and interaction violations, and achieves the highest observed completion rate over supported terrain families. Overall, TERRA provides a practical route from scene-less motion data to muscle-actuated locomotion over diverse non-flat terrain. Project website: https://cnai.epfl.ch/terra/
The Brief
2:13A short podcast summary of the paper. Written and voiced by AI.
Transcript
TERRA. Terrain-Aware Reconstruction, Retargeting and Control for Musculoskeletal Locomotion. By Merkourios Simos and three others, from EPFL. Published on the twenty-ninth of September. Its code repository is public but slight: one source file.
Muscle-actuated agents, built from musculoskeletal models and trained with reinforcement learning, can now reproduce increasingly complex human motions. Almost all of that has happened on flat ground. The abstract gives two reasons. Motion datasets rarely come with aligned terrain geometry, so a recorded person may be moving over ground that the data never describes. And even when the terrain is known, retargeting those interactions onto a complex musculoskeletal body is hard. A body made of muscles and tendons cannot simply be posed wherever the recorded skeleton happened to be.
TERRA is an end-to-end pipeline that tackles both halves at once: it recovers the ground a motion must have been performed on, and then retargets and controls the musculoskeletal body over that ground. Its starting point is deliberately thin. It works from kinematic trajectories alone, with no scene captured alongside them.
Reconstruction draws on three kinds of evidence. Terrain priors say what plausible ground looks like. Estimated contacts say where the body touched something. Negative free-space evidence works the other way round, marking space where support is absent. Together these recover the support geometry that matters for the task. Retargeting then respects three families of constraint: anatomical constraints, tendon continuity, and contact. The motion-terrain pairs that come out of this, drawn from five datasets, are used to train a single muscle-actuated control policy on nine point four hours of diverse locomotion.
The evaluation spans reconstruction, retargeting, and held-out tracking benchmarks. Across them, TERRA improves terrain accuracy and sharply reduces both anatomical violations and interaction violations. On tracking, it reaches the highest completion rate observed over the terrain families it supports.
Read as a whole, the contribution is a route rather than a single model. Scene-less motion data goes in, terrain is inferred from how the body moved, and a muscle-driven policy comes out that walks over ground that is not flat. The scope is honest about itself: the reported success is over the supported terrain families.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
People climb, sit and step down all the time on recordings. The ground they used was never recorded.
You want a simulated body driven by hundreds of muscles to walk up stairs, down ramps and onto chairs, the way people in motion-capture recordings do.
The road ahead
- 1 · The problem ~7 min Motion recordings show a body climbing stairs and ramps, but they rarely include the ground, and a muscle body cannot copy the poses directly.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min TERRA reads the ground out of the motion itself, then fits the motion to the muscle body under body and contact constraints to train one policy.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min TERRA runs from bare skeleton motion to rebuilt terrain, then a constrained fit onto the muscle body, then one policy trained on all the pairs.
- 6 · The experiment ~6 min You’ll guess: On held-out tracking over its supported terrain families, how did the policy trained on TERRA motions compare on completion rate with the other retargeting methods?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.