kinder-bench/kinder-openpi-checkpoints

Pi0.5 Fine-tuned Checkpoints for KinDER Benchmark

Fine-tuned Pi0.5 (Physical Intelligence) checkpoints for the KinDER benchmark tasks, trained using the kinder-openpi framework.

Model Description

These checkpoints are fine-tuned versions of Pi0.5, a Vision-Language-Action (VLA) model, adapted for various robotic manipulation tasks in the KinDER benchmark. The models use chain-of-thought (CoT) reasoning for improved task performance.

Available Checkpoints

3D Environment Tasks

Task	Description
basemotion3d	Basic 3D motion primitives
shelf_3d	Shelf manipulation in 3D
sweep_3d	Sweeping task in 3D environment
transport3d	Object transport in 3D

2D Environment Tasks

Task	Description
motion2d	Basic 2D motion primitives
stickbutton2d	Button pressing with stick in 2D
dynobstruction2d	Dynamic obstruction avoidance in 2D
dynpushpullhook2d	Push/pull with hook in dynamic 2D environment

Usage

Installation

Clone the kinder-openpi repository with submodules:

git clone --recurse-submodules git@github.com/Princeton-Robot-Planning-and-Learning/kinder-openpi/
cd kinder-openpi

Install dependencies using uv:

GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

Inference

Step 1: Launch the Policy Server

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_kinder_finetune \
  --policy.dir=checkpoints/<task_name>/<step>

Replace <task_name> with one of the available tasks (e.g., basemotion3d, motion2d) and <step> with the checkpoint step (e.g., 15000).

Step 2: Run Evaluation

Install the client:

pip install openpi_client tyro

Run evaluation:

# For 3D environments
python scripts/eval.py --use_overview_image --open-loop-horizon=8

# For 2D environments
python scripts/eval.py --no-use_overview_image

Training Details

• Base Model: Pi0.5 (Physical Intelligence)
• Training Framework: kinder-openpi
• Training Infrastructure: Multi-host TPU with FSDP sharding
• Data Format: RLDS (TensorFlow Datasets)

---

Citation

If you use these models, please cite the paper: KinDER: A Physical Reasoning Benchmark for Robot Learning and Planning:

@inproceedings{huang2026kinder,
  title     = {KinDER: A Physical Reasoning Benchmark for Robot Learning and Planning},
  author    = {Huang, Yixuan and Li, Bowen and Saxena, Vaibhav and Liang, Yichao and Mishra, Utkarsh and Ji, Liang and Zha, Lihan and Wu, Jimmy and Kumar, Nishanth and Scherer, Sebastian and Xu, Danfei and Silver, Tom},
  booktitle = {Robotics: Science and Systems (RSS)},
  year      = {2026}
}