OpenArm RL best policy
Best PPO teacher policy chain from the OpenArm MuJoCo cube-in-box campaign.
This is not a single monolithic checkpoint. The best teacher is a chained controller:
- Grasp leg:
checkpoints/grasp_hover_v3_vm_ppo_1000000_steps.zip
- Place leg:
checkpoints/place_fixed_v5_ppo_1000000_steps.zip
- Runtime gate/controller:
code/eval_chained_gated.py
Final validated numbers, from BEST_POLICY.md:
- In-box: 43/50 = 86.0%
- Gentle: 42/50 = 84.0%
- Eval seeds: 1000-1049
- Handover: switch after 10 consecutive grasped/lifted steps (
STREAK=10, ZTHR=0.52)
- Release gate: keep gripper closed until cube is above box footprint and low horizontal speed
See BEST_POLICY.md for exact reproduction command, known failures, and history.
Source working tree at upload time: /home/nvidia/.openclaw/workspace/projects/openarmmujoco.
