Dexmate Vega 1 Pro π0.5 — blue bird task
This private repository contains LeRobot 0.6.0 inference artifacts for a π0.5 policy fine-tuned on a single-task Dexmate Vega 1 Pro mobile-manipulation dataset.
The task instruction used during recording and training is:
put the blue bird on the meeting desk
Immutable training inputs
- Dataset:
Kasra99/dexmate_blue_bird - Dataset commit:
72a97b1a916699c17177e311463729d757f3119c - Base policy:
lerobot/pi05_base - Base-policy commit:
7de663972b7817d2c4cf2d84c821153dfea772e9 - Software: LeRobot 0.6.0, Python 3.12, PyTorch 2.11.0
The LeRobot v3 dataset has 44 episodes and 67,219 frames at 30 FPS. Training used 39 episodes, with 5 episodes held out for offline validation. It contains one 640×480 RGB head camera, a 27-value state, and a 27-value action.
Candidate revisions
Each tag points to a complete pretrained_model artifact at that training
checkpoint. The repository's main branch ends at the 30,000-step artifact.
| Tag | Updates | Held-out loss |
|---|---|---|
step-005000 | 5,000 | 0.1035 |
step-015000 | 15,000 | 0.1042 |
step-030000 | 30,000 | 0.1258 |
The 5,000-step checkpoint has the lowest offline loss, but this small
single-task holdout is not sufficient to select a deployment checkpoint.
Compare the candidates with short, guarded robot trials and pin deployment to
the chosen commit SHA rather than mutable main.
Policy configuration
- bfloat16 π0.5 policy
- 50-step action chunks
- 10 denoising/inference steps
- 27 physical action values
- 27-value input state, padded internally to 32
- quantile normalization for state and action
- expert-only fine-tuning
- relative-action conversion disabled
- one real head camera; pretrained wrist-camera slots are absent/masked
Use every file in the revision. In particular, the serialized LeRobot
preprocessor and postprocessor contain the camera rename and quantile
normalization state. Loading only model.safetensors is incorrect.
The preprocessor accepts the original dataset camera key
observation.images.head and renames it to
observation.images.base_0_rgb. It also references the tokenizer
google/paligemma-3b-pt-224, which must be available in the Hugging Face cache
for offline deployment.
Observation and action layout
The 27 values are ordered as follows:
- torso joints 1–3
- head joints 1–3
- left-arm joints 1–7
- right-arm joints 1–7
- left-hand open/close and thumb-opposition ratios
- right-hand open/close and thumb-opposition ratios
- base
vx,vy, andwz
For observations, the first 20 values are measured positions in radians, the
four hand values are measured ratios, and the final three values are measured
base velocities. For actions, the first 20 values are absolute joint targets
in radians, the four hand values are absolute ratios in [0, 1], and the last
three values are base velocity commands in m/s, m/s, and rad/s. The ROS base
frame convention is +x forward, +y left, and +z up.
Do not interpret the joint targets as deltas or velocities, and do not integrate the base commands into poses.
Download
Authenticate with a read-only token, then download an immutable candidate:
hf download Kasra99/pi05-dexmate-blue-bird \
--revision step-005000 \
--local-dir models/pi05-dexmate-blue-bird
For production, replace the human-readable tag with the resolved commit SHA. Use a LeRobot 0.6.0-compatible environment and load the downloaded directory as the policy path.
Scope and safety
This repository is a policy artifact, not a complete Dexmate controller. A deployment still requires the exact recorder-compatible state construction, RGB camera conversion, Dexmate action adapter, hard joint/base limits, freshness watchdogs, E-stop integration, and a safe chunk-execution strategy.
This policy has not yet been validated for autonomous real-robot success. Begin with offline forward passes and observe-only ROS replay. Use reduced limits, a ready E-stop, and explicit arming for any later physical trial.