Lab initiative · active

MONSTR

Humanoid foundation policy. End-to-end from physics-conditioned BC pretrain through AMP fine-tuning to deployment-grade rollouts.

What it is

A 56-joint humanoid foundation policy trained against the AMASS motion corpus with mimic-PD action targets. The foundation pretrain runs in a physics simulator; downstream deployment fine-tunes per-task while preserving the foundation's motion priors.

Status

v11 line in active iteration. PD-target action diagnostic identified head-index mapping mismatch between SMPL 24-body and SMPLX 56-body configurations as the root cause of the v11.0a-q plateau (RCA 2026-05-03).

Tracker

Active state: read /etc/monstr/flywheel.env on the training host. Result snapshots under results/. Diagnostic harness atresearch/monstr/bc_pretrain/diag_pd_action_validity.py.