pi0.5 YAM red-cap full fine-tune — step 7,500

This repository contains a full-model fine-tune of Physical Intelligence's official pi0.5 base checkpoint for a single-active-arm YAM/GELLO task: pick up a red cap and place it in a black box.

This is an OpenPI JAX/Orbax checkpoint. It is not a Transformers checkpoint and cannot be loaded with AutoModel.from_pretrained.

Model and interface

• Base: official gs://openpi-assets/checkpoints/pi05_base/params
• Training method: full fine-tuning; no LoRA and no frozen model weights
• Checkpoint: training step 7,500
• OpenPI source: npow/openpi, feat/yam-redcap-finetune
• OpenPI base commit: 215abfb217dbac7d5f1273282331b9b1866c0479
• YAM fine-tuning pipeline commit: 8539331ec241dfa3b232fc792115548348567bc5
• Model tensor width: 32 state/action dimensions, with OpenPI padding the 7 physical dimensions
• Robot-facing state/action: 7 values (6 arm joints + 1 gripper)
• Action chunk: 15 x 7, representing one second at 15 Hz
• Cameras: three RGB views, consistently mapped as middle -> base_0_rgb, left -> left_wrist_0_rgb, and right -> right_wrist_0_rgb
• Native YAM and Servo gripper convention: 0 = closed, 1 = open

The YAM transform flips the gripper to pi0.5's model-facing convention before normalization and flips it back on output. Joint channels 0–5 are trained as deltas; the gripper channel remains absolute.

The seven-dimensional policy interface is side-neutral. The deployment binding selects whether those values control the physical left or right arm; the inactive arm must be held by the robot controller.

Training data

The source collection is public: npow/yam-gello-red-cap-100.

It contains 100 demonstrations and three synchronized camera streams. The training conversion resampled trajectory timestamps to 15 Hz, selected the active left-arm dimensions, and used a deterministic episode split:

• Training: 90 episodes / 17,564 frames
• Held out: 10 episodes / 1,993 frames
• Held-out source episode IDs: 6, 16, 26, 36, 46, 56, 66, 76, 86, 96
• Prompt: pick up the red cap and place it in the black box

The exact converter, transforms, and training configuration are in the YAM example on the training branch.

Training configuration

• Model: Pi0Config(pi05=True, action_dim=32, action_horizon=15)
• Batch size: 32
• LR: cosine decay, 500 warm-up steps, 2.5e-5 peak, 2.5e-6 final at 10,000 steps
• EMA: disabled
• Seed: 42
• Initialization: official pi0.5 base only

Normalization statistics from the converted training split are bundled under assets/.

Download and serve

The checkpoint publishes openpi_config.json, exported from the training configuration and processor at commit 8539331ec241dfa3b232fc792115548348567bc5 linked above. It declares middle, left, right, and state as caller inputs and a 15 x 7 action chunk. The processor maps the three cameras to OpenPI's internal image slots and preserves the training-time state/action transforms. Those internal slot names are not caller inputs.

servo model inspect hf://npow/pi05-yam-red-cap-full-7500

The seven-dimensional interface does not choose a physical arm. The robot application supplies the observations and applies the returned actions to its intended arm.

Limitations and safety

This checkpoint is specialized to one tabletop task, one robot setup, one camera arrangement, and a small demonstration set. It should not be treated as a general-purpose manipulation policy. Camera role/order, state order, action rate, normalization statistics, and gripper semantics must match training.

Robot execution requires independent joint/rate/workspace limits, timeout-to-hold behavior, an inactive-arm hold controller, reduced-speed first rollouts, and a reachable emergency stop. The model itself is not a safety controller.