YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
pi05_rm65_left_4tasks_274 (v1)
pi0.5 fine-tuned on the RealMan RM65 LEFT arm: pick up the leftmost object of the named kind.
4 objects, 274 episodes (2026-10-16 delivery, HF LGG100/tmp; merged then trimmed: rm65/left_4tasks_274_v1_trim, 73,109 frames @ 30 fps).
- openpi config:
pi05_rm65_left_4tasks_274(SigLIP full + LLM LoRA + full action expert, action_horizon 32, EMA 0.99, batch 32, 20k steps) - state / action: 7-dim absolute joints
left_q1..q6(rad) +left_gripper(0 closed .. 100 open) - cameras:
observation/top,observation/wrist - checkpoints:
12000/,16000/,19999/β each hasparams/andassets/rm65/left_4tasks_274_v1_trim/norm_stats.json(optimizertrain_state/omitted). Last logged: Step 19900: grad_norm=0.0428, loss=0.0016, param_norm=1803.7018
Prompts (use verbatim):
- Pick up the leftmost yellow potato chip package in the top camera image with the left arm.
- Pick up the leftmost cylindrical potato chip can with red and green packaging in the top camera image using the left arm.
- Pick up the leftmost purple cylindrical potato chip can in the top camera image using the left arm.
- Pick up the leftmost container of dark-colored cola in the top camera image using the left arm.
Serve:
uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_rm65_left_4tasks_274 --policy.dir=<path>/19999
PyTorch
pytorch/19999/ is step 19999 converted with examples/convert_jax_model_to_pytorch.py (LoRA adapters merged into
the base weights, bfloat16): model.safetensors, config.json, and the same assets/.../norm_stats.json.
Checked against the JAX checkpoint on 48 frames (all 4 tasks) with identical noise: mean |Ξ| 0.003 rad on joints
(max 0.025), and the same error against ground-truth actions (joint MAE 0.0258 vs 0.0255 rad).
openpi detects the PyTorch weights by model.safetensors, so serving is the same command with the PyTorch dir
(needs transformers==4.53.2 with openpi's transformers_replace patch):
uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_rm65_left_4tasks_274 --policy.dir=<path>/pytorch/19999