Bimanual Manipulation Training Data

Claru operates ALOHA-compatible bimanual collection stations with hardware-synchronized dual leader-follower arms, wrist-mounted cameras on each arm, and 2-3 third-person cameras covering the full bimanual workspace. All sensor streams share a hardware clock signal guaranteeing sub-millisecond inter-arm synchronization, verified automatically during post-collection quality checks. Our operators are specifically selected and trained for bimanual coordination quality, completing a qualification protocol that assesses arm-to-arm timing, role consistency, and recovery behavior before production collection begins. Strict 30-minute session limits with mandatory breaks prevent fatigue-induced coordination degradation. We deliver bimanual datasets with per-arm action labels, coordination mode annotations, synchronization event timestamps, and subtask phase segmentation in RLDS or HDF5 format. Collection throughput reaches 200-400 quality demonstrations per station per day for structured bimanual tasks, enabling rapid dataset scaling for ACT, Diffusion Policy, and VLA training pipelines.

Bimanual manipulation enables tasks that are physically impossible with a single arm: holding a pot steady while stirring, stabilizing a board while screwing in a bolt, folding a towel where both ends must move simultaneously, or peeling a banana where one hand grips while the other executes a precise peel motion. These tasks require coordinated planning across two kinematic chains totaling 14+ degrees of freedom (7-DoF per arm), with the added complexity that each arm's optimal action depends on the other arm's current state and near-future trajectory.

The ALOHA project (Zhao et al., 2023) transformed the field by demonstrating that bimanual policies can be learned from remarkably small demonstration sets. Using Action Chunking with Transformers (ACT), ALOHA achieved 80-96% success rates on tasks including transferring a ball between cups (96%), zipping a bag (76%), and cooking shrimp (40 out of 50 trials) from just 50 teleoperated demonstrations per task. The key insight was action chunking: predicting sequences of 50-100 future actions at once, which naturally captures the temporal coordination between arms without explicit synchronization constraints.

The coordination challenge is what makes bimanual data uniquely valuable and difficult to generate synthetically. Two-armed manipulation exhibits distinct coordination modes — symmetric (both arms do the same thing, like lifting a box), leader-follower (one arm positions while the other works), alternating (arms take turns, like hand-over-hand rope pulling), and independent (arms do different tasks in different workspace regions). Pure simulation struggles to generate the natural coordination patterns humans use because these patterns emerge from embodied experience rather than optimization. Real teleoperated demonstrations remain the primary data source for bimanual policies.

Mobile ALOHA (Fu et al., 2024) extended bimanual manipulation to a mobile base, enabling whole-body coordination tasks like opening a two-handled cabinet while walking forward, pushing furniture, and elevator operation. These mobile bimanual tasks added a further 2-DoF base (linear and angular velocity) to the 14-DoF arm action space, creating a 16-DoF coordination problem. The dataset requirements grew accordingly: Mobile ALOHA needed 50-100 demonstrations per task but also required demonstrations of recovery behaviors when the base position drifted during dual-arm manipulation.

ALOHA Unleashed (Zhao et al., 2024) pushed bimanual manipulation to a new level of dexterity. By scaling the ACT architecture with a larger Transformer backbone and training on 500-1,000 demonstrations per task (compared to ALOHA's original 50), ALOHA Unleashed achieved consistent success on tasks previously considered too difficult: tying shoelaces (78% success), hanging a shirt on a hanger (85%), and repairing a velcro strap (91%). The critical finding was that dexterous bimanual tasks have a data efficiency threshold — below ~200 demonstrations, success rates are near-zero, but performance jumps sharply between 200-500 demonstrations as the policy captures the full distribution of coordination strategies.

DexCap (Wang et al., 2024) addressed the teleoperation bottleneck for bimanual data collection by using hand motion capture to retarget human bimanual manipulation directly to robot arms. The operator wears motion capture gloves and performs the task naturally, and DexCap solves inverse kinematics to generate feasible robot joint trajectories. On 6 bimanual tasks (pouring, scooping, handover, folding, stirring, stacking), DexCap demonstrations trained ACT policies to 72% average success — lower than direct teleoperation (85%) but with 3x faster collection throughput because operators do not need to learn the teleoperation interface.

Bi-VLA (Wu et al., 2025) explored bimanual manipulation through the lens of vision-language-action models. By fine-tuning a pretrained VLA (OpenVLA) on 2,000 bimanual demonstrations spanning 12 task types, Bi-VLA achieved 75% zero-shot success on held-out bimanual tasks given only language instructions. The model learned to infer coordination modes from task descriptions — 'hold the bowl and stir' triggers leader-follower coordination, while 'lift the box' triggers symmetric coordination. This result suggests that language-conditioned bimanual policies can generalize across tasks with sufficient data diversity.

For industrial bimanual applications, the ALOHA 2 system (Aldaco et al., 2024) demonstrated that production-quality bimanual data can be collected at scale with a redesigned teleoperation setup costing under $20K per station. ALOHA 2 improved data collection ergonomics with a reconfigurable leader arm mounting, reduced operator fatigue by 40% compared to the original ALOHA design, and standardized the recording pipeline to produce RLDS-formatted datasets compatible with the Open X-Embodiment ecosystem.

Bimanual teleoperation requires specialized hardware that preserves the operator's natural two-handed coordination. The ALOHA system uses a pair of leader arms that the operator physically manipulates in free space, with follower robot arms mirroring the motion at 50 Hz. This leader-follower approach provides the most natural bimanual control because the operator uses their native hand coordination — the same neural pathways that make everyday bimanual tasks effortless. However, the leader arms must be precisely calibrated so that the mapping between leader and follower workspace is intuitive, typically requiring 30-60 minutes of setup per session.

Temporal synchronization between arms is the single most critical quality factor in bimanual data. The two arms' proprioceptive streams, action labels, and camera frames must be aligned within 1 ms (ideally hardware-synchronized via a shared clock signal). Even 5 ms of desynchronization between arms creates coordination artifacts that degrade policy performance by 15-25% on tasks requiring tight timing like handovers and simultaneous grasps. All cameras and joint encoders should be triggered from the same hardware clock, and timestamps should be validated during quality checking.

Operator management is uniquely important for bimanual collection. Controlling two arms simultaneously is cognitively demanding — dual-task interference causes coordination quality to degrade measurably after 30-45 minutes of continuous teleoperation. Sessions should be limited to 30 minutes with mandatory 10-minute breaks. Operator qualification should assess bimanual coordination quality specifically: smooth, natural arm-to-arm timing; consistent role assignment (same arm leads consistently for asymmetric tasks); and graceful recovery from errors rather than abrupt resets.

Each bimanual demonstration should be annotated with the coordination mode (symmetric, leader-follower, alternating, or independent), synchronization events (handover moments, simultaneous grasp/release), per-arm task role labels (stabilizer vs. manipulator), and contact events between the two arms or between each arm and the workspace. For long-horizon bimanual tasks, subtask boundary annotations enable hierarchical policy training where a high-level planner selects bimanual primitives and low-level controllers execute them.

Claru operates ALOHA-compatible bimanual collection stations with hardware-synchronized dual leader-follower arms, each equipped with wrist-mounted cameras and 4+ third-person cameras covering the full bimanual workspace. All sensor streams are triggered from a shared hardware clock, guaranteeing sub-millisecond inter-arm synchronization that is verified automatically during post-collection quality checks. Our stations support both the original ALOHA (ViperX 300) and the ALOHA 2 form factor, with quick-swap end-effectors for different gripper configurations.

Our operators are specifically selected and trained for bimanual coordination quality. Each operator completes a bimanual qualification protocol that assesses arm-to-arm timing smoothness, role assignment consistency, and recovery behavior quality. Operators work in strict 30-minute collection sessions with mandatory breaks, and per-session coordination quality metrics are tracked to detect fatigue-induced degradation before it contaminates the dataset.

Claru delivers bimanual datasets with per-arm action labels, coordination mode annotations, synchronization event timestamps, task role assignments, and subtask phase segmentation. Data is formatted for direct ingestion by ACT, Diffusion Policy, and VLA architectures in RLDS or HDF5 format. For clients building multi-task bimanual systems, we provide cross-task data packages with consistent annotation schemas that enable training a single policy across diverse bimanual tasks. Our collection throughput is 200-400 quality demonstrations per station per day for structured bimanual tasks.