What's the current bottleneck in scaling manipulation training data collection?

[freya.sokolov]

This came up in a Discord I'm in and I wanted a more permanent place to discuss it. Object occlusion by the robot's own hand during the final approach to a grasp is a common and annoying perception problem - the closer the hand gets to a good grasp position, the more it blocks the camera's view of exactly what it's about to grab. Figure's fourth-generation Dexterous Hand (on Figure 02/03) reportedly offers 16 degrees of freedom per hand with sensors integrated into each finger, aimed at fine force control tasks like handling small electronic components without crushing them. Anyone want to poke holes in this?

[scott21]

Just to be precise about one thing: Payload-to-hand-weight ratio varies a lot across current dexterous hands, and it's a meaningful tradeoff - more DoF and finer sensing generally means more actuators and mass in the hand itself, which eats into the arm's usable payload budget.

[jlefebvre]

Short answer: Imitation learning from human demonstration video (without robot teleoperation data) is an appealing way to scale up training data cheaply, but it runs into the embodiment gap - human hand kinematics and force profiles don't map directly onto a robot hand's very different mechanism. In-hand reorientation (repositioning a grasped object without setting it down) is one of the more advanced manipulation skills, requiring either a highly dexterous hand with enough DoF or clever use of gravity and controlled slipping - it's an active research area rather than a solved problem.

[robertmiller]

I dealt with almost this exact situation. Bimanual manipulation - two arms coordinating on one task - is harder than it looks mostly because of the added degrees of freedom and the timing/force coordination required; a lot of 'two-handed' demos are actually closer to two independent single-hand tasks done in sequence. A lot of the manipulation shown in production demos still leans heavily on teleoperation, particularly for anything involving fine force control or novel objects - autonomous grasp success rates on genuinely unstructured, previously-unseen clutter are still well below what teleoperation can achieve. Reminds me a bit of the early drone hobbyist scene, honestly.

[garcia51]

@robertmiller This matches what I've seen too. Cable routing through a wrist joint with multiple degrees of freedom is a genuinely tricky mechanical design problem - tendons and wiring both need enough slack to avoid binding through the full range of motion without tangling or fraying over thousands of cycles. Underactuated hands (fewer actuators than joints, using mechanical coupling to shape the grasp) are a reasonable engineering compromise for robust power grasps on a budget, but they generally can't do fine in-hand manipulation the way a fully actuated hand can.

[kwilliams]

@garcia51 Worth being a little skeptical of the marketing angle here. Compliant wrists that absorb impact during a bad approach or misjudged contact reduce mechanical stress on the whole arm, which matters a lot for long-term reliability even though it's a less visible feature than the hand itself. Wet, oily, or otherwise low-friction objects are still a genuine edge case for most current hands, since tactile sensing and grasp-force controllers are typically tuned and validated on dry, higher-friction test objects.

[young56]

This is exactly the kind of context I was looking for. Grasp planning for deformable or non-rigid objects (bags, cables, cloth) remains one of the genuinely unsolved problems in manipulation - rigid-body grasp models simply don't capture how the object will behave once contact starts. Makes me wonder how this looks in another five years.

[jlefebvre]

@young56 Ran into exactly this myself. Tendon-driven fingers let you move the heavier actuators back into the palm or forearm, keeping the fingers themselves light and fast, but they introduce cable routing, tensioning, and long-term wear problems that direct-actuated fingers don't have. Vision-based grasp confidence estimation (predicting success before attempting a grasp) and tactile-based confirmation (confirming after contact) are complementary rather than competing - vision helps you choose a grasp, tactile tells you if it actually worked.

[charlesbianchi]

@jlefebvre Small correction on one detail: Palm sensing gets less attention than fingertip sensing, but a lot of power grasps (holding a box, a tool handle) rely more on palm and lateral finger contact than fingertip contact, so under-sensing the palm can leave a real blind spot in grasp confidence.

[ethan.lewis5]

@charlesbianchi Yeah, this tracks with what I've read as well. Bimanual manipulation - two arms coordinating on one task - is harder than it looks mostly because of the added degrees of freedom and the timing/force coordination required; a lot of 'two-handed' demos are actually closer to two independent single-hand tasks done in sequence. Cable routing through a wrist joint with multiple degrees of freedom is a genuinely tricky mechanical design problem - tendons and wiring both need enough slack to avoid binding through the full range of motion without tangling or fraying over thousands of cycles. Kind of makes me think about how different this all looked even three years ago.