Bimanual Manipulation Datasets

Curated open-source and custom bimanual manipulation datasets for imitation learning, VLA fine-tuning, and dual-arm policy research. Two-arm coordination is one of the hardest open problems in robot learning — these datasets provide the demonstrations to tackle it.

Key Open-Source Bimanual Datasets

Dataset	Episodes	Robot	Format
ALOHA	~50 per task, ~10 tasks	2x ViperX-300	HDF5, LeRobot
Mobile ALOHA	50+ per task	2x ViperX-300 + wheeled base	HDF5, LeRobot
ALOHA 2	Varies	Google ALOHA 2 cells	RLDS
DROID (bimanual split)	Subset of 76K	Franka + assorted	RLDS, HDF5
RH20T	110K+ episodes	Dual Franka Panda	HDF5

What Makes Bimanual Data Special

Bimanual datasets record two synchronized action streams — typically 14 DoF of joint positions plus two gripper states. Coordination timing between the left and right arms is critical: a 50 ms desynchronization can turn a successful handoff into a dropped object. This means bimanual datasets require stricter time synchronization, higher collection frequency (usually 50 Hz), and operators trained specifically in dual-arm coordination.

Common bimanual tasks include: object handoffs, box packing, lid opening and pouring, threading and tying, two-handed assembly, and cooperative lifting of large or deformable objects.

Custom Bimanual Data Collection

RCSV operates Mobile ALOHA and OpenArm bimanual collection stations in our San Francisco lab. We collect custom bimanual datasets with leader-follower teleoperation, deliver in your target format, and provide full QA with per-episode quality scores.

Need a bimanual platform? The Mobile ALOHA is available for purchase or lease.