arrow-hf/pi0-robotwin-5task-eval

pi0 RoboTwin 5-task eval — Heisen0928/pi0_robotwin, VLA vs VLA+MPC

Success-rate evaluation of the open-source Heisen0928/pi0_robotwin π0 checkpoint on 5 RoboTwin tasks, 50 seeds each (seed 100000–100049), run through the vla_framework eval pipeline. Two modes: raw policy (passthrough) and with the reactive MPC corrector.

TL;DR

task	passthrough	+MPC	Δ
place_container_plate	29/50 (58%)	23/50 (46%)	−6
place_empty_cup	26/50 (52%)	35/50 (70%)	+9
move_can_pot	21/50 (42%)	13/50 (26%)	−8
place_bread_skillet	11/50 (22%)	11/50 (22%)	0
move_pillbottle_pad	10/50 (20%)	8/50 (16%)	−2
TOTAL	97/250 (38.8%)	90/250 (36.0%)	−7 (−2.8pp)

The reactive MPC is net-negative here (−2.8pp). Only place_empty_cup benefits (+9, a "close-but-incomplete" task where correction recovers grasps); on tasks the raw policy already handles (plate, can_pot) the MPC perturbations (pos_std=0.015, rot_std=0.03, gripper-flip) disrupt otherwise-successful trajectories. MPC helps where the policy is marginal, hurts where it is already competent. The MPC defaults are tuned for SmolVLA-style reactive correction; π0 (continuous gripper, FK→ee path) needs a higher intervention threshold / smaller perturbation.

Heisen0928/pi0_robotwin is a multi-task π0 (no per-task finetune), hence the modest 38.8% raw rate; plate/cup strongest (>50%), pillbottle/bread weakest (~20%).

Setup (reproducible)

• Checkpoint: Heisen0928/pi0_robotwin (JAX/orbax, step 30000) → converted to PyTorch with openpi's convert_jax_model_to_pytorch.py (--config_name pi0_aloha, architecturally identical: action_dim 32, action_horizon 50, gemma_2b + gemma_300m, bf16). norm_stats from 30000/assets/robotwin_50_clean/ placed at physical-intelligence/robotwin/norm_stats.json.
• Backend: in-process Embodied-AI VLAPredictor (model_type: pi0, config_name=pi0_robotwin), JOINT(qpos,14D) policy bridged to the ee eval path via RoboTwin FK (fk: joint_to_ee).
• Poses: pose_source: robotwin_gt (GT, no poselib — isolates the policy).
• Env: conda vla_unified; MUJOCO_GL=egl, CUDA_VISIBLE_DEVICES=1, EMBODIED_SKIP_TRANSFORMERS_CHECK=1. subprocess-per-episode (SAPIEN leak).
• Per-episode: --num_episodes 1 --max_steps 300.

Configs in configs/ (pi0_robotwin_pass.local.yaml, pi0_robotwin_mpc.local.yaml).

Contents

• results_per_seed.csv — per task × seed, passthrough & mpc outcome (SUCCESS/FAIL), 250 rows.
• SUMMARY_pass.txt, SUMMARY_mpc.txt — per-task aggregates.
• configs/ — the exact eval configs (task overridden per run via --task).