A robot releases a snap fit. Another robot has a different gripper, different tactile sensors, and a different force range. Replaying the first robot’s joint trajectory is unlikely to reproduce the task. What, if anything, should carry over?

One proposal from my recent research planning represents a skill through contact transitions: establish contact, load the mechanism, detect the release, then move into a terminal state. The representation would describe relations among the tool, object, and task rather than pretending the two bodies share actuators.

This is a research proposal. I do not have a demonstrated cross-body transfer result. The difficult part is specifying an abstraction that preserves useful information without hiding the calibration each new body still needs.

A contact state is partly hidden

A force spike might indicate the latch has loaded, an unintended collision has occurred, or a sensor has drifted. Geometry, motion history, and the distribution of force across the contact surface can distinguish some of these possibilities.

The controller needs a belief about the contact state, not just a label attached to one tactile frame. It also needs a local model of how its own actions change that state. The same object-relative objective can require different motions from a parallel gripper and a compliant hand.

A proposed skill representation would therefore include a contact-state estimate, transition conditions, a terminal condition, and uncertainty about the local dynamics. The target body supplies its own embodiment-aware controller. Calibration probes estimate the parts that cannot transfer.

The calibration bill is the experiment

A transfer method can appear impressive if the target robot receives extensive per-task tuning that the baseline does not. Counting only the final demonstration conceals the main expense.

For snap-fit extraction, begin with a source demonstration and held-out part variants. Give the target a limited number of probes, then measure successful extraction, damage, force-limit violations, human attention, and total adaptation time. Compare with target-only learning and a specialist controller using the same sensing and calibration budget.

The key comparison is how quickly the transferred structure becomes useful. If the target needs as much identification and supervision as it would from scratch, the shared representation may be descriptive without being operationally valuable.

What should remain invariant?

“Release the latch” is too abstract to control the robot. “Move motor three by this amount” is too specific to transfer. The useful intermediate description might include object-relative motion, a contact constraint, and an event such as a drop in stored elastic load.

Even those events need task-specific checks. A force drop can signal release or breakage. A terminal visual pose can hide damage. Acceptance should be tied to a physical task contract rather than a classifier trained to recognize the source robot’s successful motions.

The representation should also expose where it stops applying. A different friction regime or a new latch geometry may require a new skill rather than a small parameter update. An honest failure detector can be more useful than broad but brittle transfer.

Beyond one additional hand

A single source and target pair can support a narrow finding. It cannot establish an embodiment-independent skill language. A stronger sequence would hold out a part, then a sensor configuration, then another hand. At each stage, ask which elements stayed reusable and which had to be relearned.

The long-term idea is a library of physical skills whose reusable content is measured by the work it saves on a new machine. The first result would be much smaller: one contact task, a second body, and a complete account of the probes needed to make the transfer work. That account is where the interesting physics would appear.

Related notes