---
title: pi05_robocasa_gr1_tabletop24
canonical_url: "https://www.modelscope.cn/models/NewGain/pi05_robocasa_gr1_tabletop24"
md_url: "https://www.modelscope.cn/models/NewGain/pi05_robocasa_gr1_tabletop24.md"
repository: NewGain/pi05_robocasa_gr1_tabletop24
last_updated: 2026-07-31
library_name:
  - jax
frameworks:
  - jax
downloads: 0
stars: 0
---

# pi05_robocasa_gr1_tabletop24

> pi05_robocasa_gr1_tabletop24 - NewGain 在 ModelScope 开源的模型。pi0.5 RoboCasa GR1 Tabletop-24 Inference Package

- **Repository**: NewGain/pi05_robocasa_gr1_tabletop24
- **Downloads**: 0
- **Stars**: 0
- **Last updated**: 2026-07-31

Source: https://www.modelscope.cn/models/NewGain/pi05_robocasa_gr1_tabletop24

---

# pi0.5 RoboCasa GR1 Tabletop-24 Inference Package

This package is inference-only. It contains model params, checkpoint assets,
the frozen GR1 observation/action adapter, the OpenPI runtime source, the
Paligemma tokenizer, and audit metadata. It intentionally excludes optimizer
and train state.

## Verify and extract

```bash
sha256sum -c pi05_robocasa_gr1_tabletop24_formal_20260730_01_step30000_inference_only.tar.sha256
mkdir -p ~/models/pi05_gr1_tabletop24_step30000
tar -xf pi05_robocasa_gr1_tabletop24_formal_20260730_01_step30000_inference_only.tar \
  -C ~/models/pi05_gr1_tabletop24_step30000 --strip-components=1
cd ~/models/pi05_gr1_tabletop24_step30000
sha256sum -c PACKAGE_MANIFEST.sha256
```

Use the archive for the predetermined step 30,000 checkpoint in formal
evaluation. Step 10,000 and 20,000 packages are recovery/intermediate
deliverables, not candidates to select using rollout results.

## Environment

Use Ubuntu 22.04, Python 3.11, a CUDA 12 driver/runtime compatible with JAX
0.5.0, and the repository's pinned dependencies:

```bash
cd openpi_runtime
uv sync
export OPENPI_DATA_HOME="$(cd .. && pwd)/openpi_cache"
export PYTHONPATH="$(pwd)/src"
export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1
```

The package was already loaded and run in a fresh CPU process before delivery.
To repeat that non-rollout gate:

```bash
JAX_PLATFORMS=cpu CUDA_VISIBLE_DEVICES="" \
  uv run python scripts/validate_gr1_portable_package.py \
  --package-root .. \
  --output ../LOCAL_FRESH_PROCESS_VALIDATION.json
```

## Policy contract

The policy input is:

- `observation/image`: one ego-view RGB image, `H x W x 3` uint8 or float.
- `observation/state`: the original 44D GR1 state.
- `prompt`: the original coarse action instruction.

The model returns `16 x 29` absolute actions in this frozen order:

1. left arm, 7D
2. right arm, 7D
3. left hand, 6D
4. right hand, 6D
5. waist, 3D

Use `openpi.policies.gr1_tabletop_policy.active_to_action_dict` to obtain:
`action.left_arm`, `action.right_arm`, `action.left_hand`,
`action.right_hand`, and `action.waist`.

Do not apply `DeltaActions`. The two wrist camera slots are empty and masked.
The shared all-24 normalization asset is under
`checkpoint/assets/robocasa_gr1_tabletop24_global/`.

## Base vs AHS

For the rebuttal comparison, freeze the same step-30,000 package, norm asset,
task list, episode seeds, simulator settings, instructions, and policy sampling
parameters for both methods.

- Base executes a fixed horizon from the declared candidate grid
  `{4,8,12,16}`.
- AHS may choose only from `{4,8,12,16}` and changes execution/replanning
  horizon, not the generated 16-step action chunk or base-policy weights.
- Do not use rollout outcomes to choose between the 10k/20k/30k checkpoints.
- Run a small fixed-manifest smoke first, then freeze the formal manifests
  before launching the 24-task evaluation.

Formal RoboCasa rollout wiring is environment-specific and must adapt simulator
observations to the policy contract above. Preserve per-replan selected horizon,
policy calls, inference time, wall time, and outcome in the result records.