TERRA-4B

The TERRA PPO checkpoint used for the primary, training-seed-0 manuscript tables.

Checkpoint

• Environment steps: 4,000,317,440.
• PPO update: 24,416 (checkpoint_24416).
• Training environment: MjxMyoFullBody.
• Training method: TERRA terrain reconstruction and motion retargeting.

The full Orbax checkpoint includes the configuration, training state, observation normalization state, optimizer state, and training metadata.

Download and visualize

Use the TERRA code at https://github.com/amathislab/terra (release/code branch).

uvx --from huggingface_hub hf auth login
export POLICY_CHECKPOINT="$HOME/terra-results/checkpoints/TERRA-4B"
uvx --from huggingface_hub hf download merc-s/TERRA-4B \
  --local-dir "$POLICY_CHECKPOINT"
export POLICY_DATASET="/absolute/path/to/materialization.json"
python scripts/terra/play_policy.py \
  --checkpoint "$POLICY_CHECKPOINT" \
  --materialization-record "$POLICY_DATASET"

Create materialization.json with terra train materialize as described in the TERRA policy guide.

Add --video-dir /absolute/path/to/videos to record MP4 files instead of opening the native MuJoCo GUI. Use --motion <identifier> to select one dataset motion.