niejunnan25/pi05-robotwin

π₀.₅ RoboTwin Checkpoints for PACE

Fifty task-specific π₀.₅ checkpoints used in PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (paper). Each task has its own model and normalization statistics. PACE is a training-free execution method: Fixed-horizon, PACE and the adaptive baselines in the paper use the same task-specific model weights. PACE calibration results and evaluation code are distributed with the companion codebase.

Contents

pi05-robotwin/
  manifest.json
  tokenizer.model
  beat_block_hammer/
    params/
    assets/robotwin/norm_stats.json
  ...49 other tasks...

Select a task using its standard RoboTwin name. A single task is approximately 12.4 GB; the complete collection is approximately 622 GB. manifest.json lists every task, its OpenPI configuration, asset files, sizes and SHA256 checksums. Model parameters retain the JAX/Orbax OCDBT format. Normalization statistics are the original task-specific statistics; do not replace them with pooled values.

Training

For each task, the pretrained π₀.₅ model was fine-tuned on 50 demo_clean demonstrations. All parameters were updated using AdamW, global batch size 32 and training seed 42. The learning rate warmed up to 2.5e-5 over 1,000 steps, then followed a cosine schedule toward 2.5e-6. These are the checkpoints saved at step 30,000, using EMA weights with decay 0.99.

The policy consumes three CHW-format uint8 camera images and a 14-dimensional dual-arm joint/gripper state. It returns a 50 × 14 chunk of absolute joint/gripper actions. The normalization and action transforms are supplied by the corresponding pi05_robotwin_<task> configuration.

Load with OpenPI-PACE

Set up OpenPI-PACE first. If you already have a working OpenPI installation, the guide explains how to add the RoboTwin configurations while reusing your workspace, environment, and cache. No training demonstrations are required for inference.

The checkpoint configurations were originally verified with source revision ba9876639cb7a4bf0c7db00ec39eeb9c8efd8b12.

Download a checkpoint

From your OpenPI workspace, with its model environment activated, download the adjust_bottle checkpoint and shared tokenizer:

huggingface-cli download niejunnan25/pi05-robotwin \
  --include "adjust_bottle/*" tokenizer.model manifest.json TERMS.md NOTICE \
  --local-dir local/pi05-robotwin

Replace adjust_bottle/* to download another task. Existing copies of the same checkpoints can be reused. Keep each task's params/ and matching assets/robotwin/norm_stats.json together.

For the first run, place the tokenizer in OpenPI's default cache. Skip this if it is already there:

mkdir -p ~/.cache/openpi/big_vision
cp local/pi05-robotwin/tokenizer.model ~/.cache/openpi/big_vision/paligemma_tokenizer.model

For a custom cache, keep using OPENPI_DATA_HOME; the tokenizer belongs at big_vision/paligemma_tokenizer.model within that directory.

Start the model server

Run the standard OpenPI server from the same workspace and model environment:

CUDA_VISIBLE_DEVICES=0 XLA_PYTHON_CLIENT_MEM_FRACTION=0.2 \
python scripts/serve_policy.py --port 8000 policy:checkpoint \
  --policy.config pi05_robotwin_adjust_bottle \
  --policy.dir local/pi05-robotwin/adjust_bottle

Follow the RoboTwin-PACE evaluation example in a separate RoboTwin Conda terminal, using the same task and port. The client selects Fixed horizons or PACE; the model server always returns a complete predicted action chunk. After evaluation, stop the server with Ctrl-C and start a fresh server for the next run.

These checkpoints cover the paper's RoboTwin2.0 task-specific models, not a single jointly trained 50-task policy or the separate LIBERO/real-robot models.

Terms and attribution

See TERMS.md and NOTICE for the upstream model and tokenizer terms. The OpenPI source-code license and the model terms are separate.