OpenPI π0.5 fine-tuned on ARX X5 single-arm real-robot data with action-chunk inpainting (RePaint-style conditional sampling) enabled. Used for VR teleoperation with packet-loss tolerance — model fills in missing VR commands by conditioning on the received past actions.

Base model: pi0.5 (cross-embodiment pretrained)
Action representation: deltapose (6-DoF delta xyz+rpy + binary-snap gripper)
Action horizon: 50 (1.67s @ 30 Hz)
past_offset: 25 (straddling chunk: past 25 + future 25)
enable_action_inpaint: True (RePaint hard-projection during sampling)
Backend: PyTorch (model.safetensors), bfloat16 inference
Train data: 415 episodes, 97 483 frames @ 30 fps
Train mask patterns: 40% tail, 25% burst, 15% iid, 20% all-known-past (simulates VR network conditions: tail drop, burst loss, random drops, ideal)
Train steps: 30 000

Inference-only — optimizer.pt (13GB) was stripped to keep the upload small.

Download

huggingface-cli download magic0/pi05-xzj-arx-deltapose-puttube-inpaint-30000 --local-dir ./checkpoints/inpaint_30000

Deploy

Server code: https://github.com/SII-ZijunX/openpi-inpaint-vla (branch inpaint-vla).

bash

git clone https://github.com/SII-ZijunX/openpi-inpaint-vla.git
cd openpi-inpaint-vla
uv venv && source .venv/bin/activate
uv pip install -e ".[pytorch]"

# IMPORTANT: must pass --enable_inpaint + --past_offset 25 for proper inpaint behavior
# (--asset_id xzj_data_0511 is required because norm_stats is under assets/xzj_data_0511/)
python scripts/inference_server_xzj_arx.py \
    --config_name pi05_xzj_arx_deltapose_puttube_0511_inpaint \
    --checkpoint_dir ./checkpoints/inpaint_30000 \
    --asset_id xzj_data_0511 \
    --enable_inpaint \
    --past_offset 25 \
    --zmq_port 4830

How inpaint works at inference

Client sends observation["past_actions_raw"] (25, 7) — training-style cumulative delta from chunk anchor, derived from per-slot VR commands.
Client sends observation["past_action_mask"] (25,) bool — True where VR for that slot was received, False where lost.
Server's _build_element quantile-normalizes past_actions and passes observed_actions + action_mask to model.sample_actions(...).
Model does RePaint conditional sampling: at each Euler step, hard-projects the known positions back to observed values; unknown positions evolve via flow matching. Result: a (50, 7) chunk with known slots == observed and unknown slots smoothly inpainted by the model.
Server slices off the past 25 slots (payload_actions[25:]), returns only the future 25 to the client for execution.

Internals

model.safetensors — PyTorch weights (~7GB), bfloat16-castable
metadata.pt — checkpoint metadata (step number, config name, etc.)
assets/xzj_data_0511/norm_stats.json — quantile normalization stats (q01/q99) used for both Normalize input and Unnormalize output. Quantile, not z-score — see use_quantile_norm=True in data config.
assets/xzj_data_0511/norm_stats_meta.json — sentinel tuple (enable_action_inpaint, past_offset, action_horizon, config_name, timestamp). Must match the train config to load successfully.

magic0/pi05-xzj-arx-deltapose-puttube-inpaint-30000

pi0.5 ARX deltapose inpaint checkpoint — puttube task (step 30000)

Download

Deploy

How inpaint works at inference

Internals

magic0/
pi05-xzj-arx-deltapose-puttube-inpaint-30000

pi0.5 ARX deltapose inpaint checkpoint — `puttube` task (step 30000)