latency-sensitive-bench/mikasa-robo

MIKASA InterceptGrabFast H1 PPO teacher

The small privileged-state PPO teacher used to collect the published InterceptGrabFast-VLA-v0 H1 rollout dataset. The task runs at 20 Hz with a 7D pd_ee_delta_pose action. The checkpoint is evaluated with deterministic mean actions.

The full setting and commands are in the MIKASA cookbook. The matching rollouts are in the dataset repository.

Files

oracle_checkpoints/InterceptGrabFast-VLA-v0/final_success_ckpt.pt
evaluation/ppo_teacher_strict_eval_tuned_final/summary.json

final_success_ckpt.pt is the PyTorch checkpoint consumed by the official MIKASA collector and evaluator. The evaluation summary records the fixed-seed 100-episode canonical evaluation. The checkpoint is 1,170,029 bytes with SHA256 3358b1a8e6d4bf88b7692e2dd1200034a9db5dd28d2c08ecafed1cbb8ebd2c40.

Training setting

• Observation: 49D privileged state.
• Action: 7D H1 end-effector delta pose.
• Reward: normalized_dense.
• Seed: 123.
• Learning rate: 1e-4.
• PPO update epochs: 8.
• Parallel environments: 1,024.
• Batch size: 61,440.
• MIKASA integration revision: cf4f96f319022f89c9d7cfbd639d19bc10ed44fb on vendor baseline 16634db18bef08128ed79346469c86fc12169aed.
• Training result: official early stop at 12,288,000 environment steps after passing the 16/16 trainer gate.

Evaluation

The canonical strict evaluation used seeds 4242424242..4242424341. The teacher passed 100/100 episodes with mean return 36.819.

After cloning the benchmark and installing its MIKASA environment:

hf download latency-sensitive-bench/mikasa-robo \
  oracle_checkpoints/InterceptGrabFast-VLA-v0/final_success_ckpt.pt \
  --local-dir outputs/mikasa/published_teacher

third_party/MIKASA-Robo/.venv/bin/python scripts/mikasa/evaluate.py \
  --policy ppo \
  --checkpoint outputs/mikasa/published_teacher/oracle_checkpoints/InterceptGrabFast-VLA-v0/final_success_ckpt.pt \
  --episodes 100 \
  --output-dir outputs/mikasa/published_teacher_eval

This checkpoint is intended as a rollout teacher and zero-delay baseline for the matching MIKASA environment. It is not the trained StarVLA student. No license metadata is asserted here.