ฯ0.5 Tool Hang โ step 30000
Full ฯ0.5 fine-tuning checkpoint for the robomimic Tool Hang simulation task.
The historical configuration name/path contains pi0_tool_hang, but this model
was trained with Pi0Config(pi05=True).
Contents and inference
30000/params contains OpenPI JAX parameters. 30000/assets contains normalization
statistics. Optimizer/training state is also retained for recovery. Use the
matching training configuration and observation transforms, not a generic LIBERO
adapter: proprioception is end-effector position (3), quaternion (4), and gripper
joint positions (2); views are sideview and robot0 wrist. Actions are native
7-dimensional OSC delta commands. Model action dimension is padded to 32, action
horizon 50; reported evaluation executes 8 actions per policy call.
Prompt: Assemble the frame and hang the tool on the frame.
Matching implementation: compress_vla/src/compress_vla/robomimic_pi0/ in
NguyenHuuChi/VLA_Quantization,
including config.py, policy_transforms.py, and evaluate.py.
Evaluation (completed 2026-10-04)
Dense 10-step inference: 54/100 test, 36/50 validation, 90/150 combined. These are separate pre-existing evaluation splits; the combined value is not a new independent test. Rollout horizon 700, control frequency 20 Hz. Performance is specific to the matching simulator/controller/observation configuration.
Whole loaded-model SHA-256:
cb1799ec56fb1f060b79349eb6955d66b89737eab889b0c2a17923a973340bb8.
Archived files were verified against remote sizes and LFS SHA-256/Git blob hashes before deleting the local step-30000 directory. This is a research checkpoint, not a validated real-robot deployment policy. Upstream model, software, and data terms still apply; no additional license grant is implied here.
- Downloads last month
- 24