MANTRA
The NONHUMAN team built an ALOHA-style bimanual manipulator using Piper arms, achieving a system 4x cheaper than Stanford's original ALOHA. I led the mechanical design and developed the teleoperation and data-collection stack. The team collected 3,000+ pick-and-place episodes (2,301 curated and available on HuggingFace) to fine-tune and test the π0.5 model.
Technologies
Overview
MANTRA (MANipulator for TRAining) is an ALOHA-style bimanual teleoperation system built using Piper robotic arms. The project was developed at NONHUMAN Lab to enable efficient data collection for imitation learning research.
The system consists of two pairs of arms: the leader arms controlled by a human operator, and the follower arms that mirror the movements in real-time. This setup allows for intuitive data collection of complex manipulation tasks.
Across all iterations, we collected 3,000+ pick-and-place episodes (2,301 curated) used to train the latest MANTRA model demonstrated in the videos. All NONHUMAN datasets are open source and available at [https://huggingface.co/NONHUMAN-RESEARCH/datasets](https://huggingface.co/NONHUMAN-RESEARCH/datasets).
Key Features
- Bimanual teleoperation with 12 DOF (6 per arm)
- Real-time motion mirroring between leader and follower arms
- Integrated camera system for visual learning
- 2,301 curated episodes (3,000+ collected) used to train the latest model shown in demos
- Open source datasets available on HuggingFace
Team
NONHUMAN Lab Team
Reflection
MANTRA represents our commitment to democratizing robotics research. By building affordable teleoperation hardware and sharing our datasets openly, we aim to accelerate progress in embodied AI and make these technologies accessible to researchers worldwide.