OpenPyRo-A1: An Open Python-based Low-Cost Bimanual Robot for Embodied AI

Framework

What we have done:

OpenPyRo-A1: Developed a low-cost dual-arm humanoid robot.
Distributed Framework: Built a Python-first control framework with VR teleoperation support.
Bi-manual Dataset: Collected manipulation episodes for skill learning.
Agentic Planning: Proposed a framework using large language and vision-language models.

Imitation Learning

▪️ ACT

To enable dual-arm manipulation, we designed tasks that required coordinated operation of both hands. Data were collected using our distributed teleoperation system, and subsequently used to train the ACT model for inference during deployment.

Cut Corn

Play Music Two Tigers

Pick Apple and Lemon Several Times

Exchange Banana

▪️ π₀

We conducted further experiments on VLA, using π₀ as the pre-trained model for fine-tuning.

Open Drawer and Pick

Pour Rice

Clean Table

Stack Cups

Teleoperated Tasks

▪️ Two-handed Collaboration

With our designed distributed robot control framework, IK control algorithm, we can smoothly control the robot to complete the dual-arm cooperative task.

Cut Apple

Unpack and Fetch

Fold Clothes

Wash Dishes

Pass the Fruit

Open the Bottle

▪️ Long-Horizon Tasks

Our robot can easily complete long-horizon tasks.

Arrange Tableware

Clean Board and Write

▪️ Precise Control

With our high frequency motor control and calibration, our robot can easily complete precise tasks.

Write with Marker Pen

Embodied AI

We present an agent framework that uses a large language model (Deepseek) and a vision-language model (InternVL) to orchestrate a library of learned policies, enabling task execution from natural spoken language instructions. We demonstrate that the system can parse the scene, select relevant skills, and successfully complete a high-level task.