Beijing AI Academy's RoboCOIN Gives Two-Armed Robots 180,000 Training Demos
A 180,000-demonstration dataset spanning 15 different two-armed robots aims to unlock cross-platform learning for bimanual manipulation tasks.
ProTakeaways
- RoboCOIN releases 180K+ bimanual demonstrations from 15 robot platforms spanning dual-arm, half-humanoid, and humanoid configurations.
- Hierarchical capability pyramid provides trajectory, segment, and frame-level annotations synchronized across every demonstration.
- Robot Trajectory Markup Language filters 35% of low-quality trajectories and yields a 22.2% relative performance gain.
- Cross-embodiment pre-training lifts success rates 75-213% on unseen ARX-X5 and Franka Panda simulator platforms.
- Full stack installs via
pip install robocoin, built on LeRobot with Apache 2.0 license. - Data mirrored on the Hugging Face RoboCOIN organization with per-task and per-embodiment subsets.
Teaching robots to use both hands the way people do has been stuck on a mundane bottleneck: there just isn't enough data. Every research group builds its bimanual setup differently, so demonstrations recorded on one robot rarely transfer to another. A team led by researchers at the Beijing Academy of Artificial Intelligence is trying to break that logjam with RoboCOIN, an open-source dataset that pools over 180,000 demonstrations collected from 15 distinct robotic platforms into a single, unified format.
The collection covers 16 scenarios across residential, commercial, and working environments, with 421 tasks organized by bimanual coordination patterns and object properties. That breadth is unusual for two-armed data, which historically has come either from tightly controlled single-platform studies or from aggregated collections dominated by single-arm robots.

Why two-armed data is scarce
Single-arm manipulation datasets have grown enormously, while two-armed platforms differ radically from each other. One robot may use parallel grippers on 6-DOF arms; another may have 7-DOF arms terminating in five-fingered dexterous hands; a humanoid might add mobile-base coordination on top of that. Kinematic conventions, coordinate frames, gripper state encodings, and camera placements all vary. The result is a fragmented ecosystem where each new lab reinvents its own data pipeline and models trained in one embodiment struggle to generalize.
RoboCOIN attacks the problem head-on by including three categories of hardware in one release: dual-arm robots such as the AgileX Cobot Magic and Galaxea R1 Lite, half-humanoid platforms like the Realman RMC-AIDA-L and Airbot MMK2, and full humanoids including the Unitree G1edu-u3. Roughly half of the demonstrations come from half-humanoids, with the remainder split between dual-arm setups and full humanoids. End effectors span both parallel jaws and dexterous multi-finger hands, and teleoperation methods range from isomorphic leader-follower arms to VR headsets, exoskeletons, and motion capture.