yeeeiii111/pi05-robotwin-icl-arx5-v4mix-20k

pi0.5 — RoboTwin ICL ARX-X5 v4mix — step 20000

Fine-tune of the Physical Intelligence pi0.5 base policy on the RoboTwin ICL benchmark for the ARX-X5 dual-arm robot, with additional paired-variant demonstrations mixed in (v4mix).

Contents

File	Purpose
model.safetensors	~7.0 GB — main model weights (bf16)
optimizer.pt	~13 GB — AdamW state; needed only to resume training
metadata.pt	Lightning bookkeeping (step, epoch, RNG state)
assets/robotwin-icl-arx-x5-v4mix/norm_stats.json	Per-dim action/state norm stats used at inference

How to load

# Requires openpi (Physical Intelligence). See
# https://github.com/Physical-Intelligence/openpi
from openpi.policies import policy_config
from openpi.training import config as pi_config

train_cfg = pi_config.get_config("pi05_robotwin_icl_arx_x5_v4mix")
policy = policy_config.create_trained_policy(
    train_cfg,
    checkpoint_dir="path/to/this/checkpoint",
)
actions = policy.infer(obs)["actions"]   # (50, 14) action chunk

Training recipe

• Base: pi0.5 (openpi pi05_base params)
• Continued from: pi05_robotwin_arx5_icl_20260702/200000 (an earlier ICL-only fine-tune)
• Config: pi05_robotwin_icl_arx_x5_v4mix
• Model: Pi0Config(action_dim=32, action_horizon=50, pi05=True, paligemma_variant='gemma_2b_lora', action_expert_variant='gemma_300m', dtype=bfloat16)
• Data:
  • Canonical: 3000 train + 1000 val episodes from robotwin-arx5-lerobot (25-task ICL split)
  • Variants: 1500 train episodes from robotwin-arx5-variants-lerobot (paired-v4 waypoint / grasp-point / arm variants)
• Optimizer: AdamW(b1=0.9, b2=0.95, eps=1e-8, wd=1e-10, clip_grad_norm=1.0)
• LR schedule: CosineDecay(warmup=200, peak_lr=3e-5, decay_steps=50000)
• Freeze filter: LLM backbone frozen except _1 (action expert) and LoRA adapters; PaliGemma trained with gemma_2b_lora
• This checkpoint: step 20000 / 50000, batch size 16 × 4 GPUs (FSDP), bfloat16 precision
• Trained on: 2026-07-06

Notes on step choice

Prior top-K noise-coverage diagnostics on the v4mix training series found that step 20k retains more multimodal coverage of variant modes (arm switch, waypoint direction, grasp point) than later steps 35k+, which progressively collapse toward the modal action. For downstream residual adapter / noise-steering experiments that rely on the base policy having non-trivial support over variant modes, 20k is a better starting point than converged 35k+.

For strict in-distribution BC accuracy on canonical demos, later checkpoints (35k / 50k) may be preferred.

Closed-loop sanity (small smoke)

On beat_block_hammer_D_wpN val split (arx-x5, SAPIEN sim), 3 seeds: 2 successes out of 2 solvable seeds (1 seed marked expert_fail by the planner and skipped). Full closed-loop success rate table has not been computed yet.

License

Model weights are derived from Physical Intelligence's pi0.5 base and inherit its Apache-2.0 license. Training data (RoboTwin) has its own license — see the RoboTwin repo.