Ask a humanoid company for a demo and you will see locomotion, box moving, and perhaps a tabletop pick. Ask the same robot to button a shirt cuff, pour a glass to a consistent level, or carry a loaded tray through a doorway, and the demo ends. That gap is not an accident of marketing. It maps precisely onto the tasks that demand fine motor control, and those tasks are underrepresented in training data by orders of magnitude relative to their deployment value.

The task families, as we capture them

On our floor, fine motor work is organised into recurring families rather than one-off requests, because policies generalise within families far better than across them.

  • Fastening. Buttons, zips, snaps, buckles, laces, hooks. Two-handed by nature: one hand stabilises material while the other executes a sub-centimetre alignment, usually against fabric that moves when touched.
  • Folding and flattening. Garments, linens, towels. Long-horizon sequences of grasp, trace, and crease where each step reshapes the object the next step depends on.
  • Pouring and decanting. Liquids and granular media into varied vessels. Continuous control against a moving fill level, with tilt rate as the difference between a pour and a spill.
  • Tray service and handover. Carrying loaded, shifting payloads, placing items without disturbing neighbours, and passing objects to a human at a natural height and moment.
  • Insertion and threading. Plugs, keys, cable ends, needle-and-slot assemblies. Millimetre tolerances that must be met through contact feel rather than vision alone.
  • Tool use in hand. Regrasping and reorienting an object within the hand, using a tool while holding it, stabilising work with the second hand.

Why these tasks resist current policies

Four properties compound. First, the tolerances sit at millimetre scale, where success and failure look nearly identical from any camera more than arm's length away. Second, the decisive variable is often force, not position: how hard to pinch a button, how firmly to hold a tray, and a vision-trained policy must infer force consequences from what it can see. Third, most fine motor work is bimanual, and coordinating two hands multiplies the control problem rather than adding to it. Fourth, these are long-horizon tasks with strict phase structure, and an error in an early phase, a bad initial grasp on a zip pull, surfaces as a failure thirty seconds later where it is hard to attribute.

And cutting across all four: much of this work happens on cloth. Fastening a cuff is simultaneously a fine motor problem and a deformable object problem, which is why we treat garment manipulation data as its own discipline and why tasks at that intersection are the scarcest data of all.

Why this is the frontier right now

The economics moved. Gross manipulation, reaching, lifting, transporting rigid objects, is increasingly covered by existing corpora and improving simulators, so it no longer differentiates anyone. The deployments the industry is actually selling, hospitality service, care support, domestic assistance, commercial laundry, are gated on exactly the families above; a service robot that cannot manage tray work or fastening is a kiosk on legs. Meanwhile the scaling evidence, including the log-linear demonstration-to-performance relationships NVIDIA researchers have reported, implies these skills are learnable with enough of the right data, and the right data barely exists publicly. Fine motor demonstrations are where data spending buys the most differentiation per episode, which is why labs concentrate their budgets here and why our capture calendar does too.

What fine motor capture must include

Recording someone doing up a button is easy. Recording it so a policy can learn from it imposes requirements that generic video capture misses:

  • Wrist-level close views with locked exposure, because the decisive geometry occupies a few square centimetres and vanishes at distance.
  • High frame rates at the contact moment, so the alignment micro-corrections that constitute the skill are actually sampled rather than blurred through.
  • Fine-grained phase labels, reach, stabilise, align, execute, verify, anchored to observable events, because long-horizon credit assignment needs the structure.
  • Graded outcomes with failure classes, since near-misses carry the gradient. A button seated on the second attempt is a more valuable episode than a clean first-try success.
  • Variation across instances, button sizes, zip stiffness, vessel shapes, tray loads, planned as a coverage matrix rather than accumulated by accident.
  • Gripper-aware demonstration protocols when the target platform is known, constraining human technique to strategies the end effector can actually execute.

That last requirement feeds a sourcing question we get on almost every scoping call: whether fine motor data should come from human hands or from teleoperation on the robot itself. The dexterity ceiling of current teleoperation rigs bites hardest precisely on these tasks, which is a large part of our honest comparison of the two approaches.

How we tier difficulty inside a family

Task cards for fine motor families define difficulty tiers along the axes that actually move failure rates. For pouring: fill level, media viscosity, vessel mouth diameter, target fill tolerance, distractor clutter. For fastening: fastener size, fabric stiffness, garment orientation, one-handed versus assisted. Programmes typically ramp through tiers, front-loading the middle of the difficulty range where learning signal is richest, then extending to the extremes once the client's policy has a foothold.

Tray service is a useful example of why tiering matters. Tier one is a rigid tray with fixed, stable items on a clear path. By tier four the tray carries an open vessel of liquid, items with different friction against the tray surface, and a doorway transit with a pause and restart. A policy trained only on tier one collapses at tier four, but a corpus captured only at tier four is too chaotic to bootstrap from. The ramp is the curriculum. It is a small piece of design, but it is the difference between a dataset that teaches a skill and one that merely documents people succeeding at it.