Does the gripper change
when to look again?

Same setting, different outcome.

We changed the physical response of the two finger motors: nominal stiffness/damping of 2,000 N/m and 100 N·s/m versus a softer setting of 800 N/m and 30 N·s/m. These motor gains change how the fingers respond to commands; their shape and material are unchanged. Each gripper setting used fresh images every command or reused visual features for eight commands. The frozen policy, action-plan interval and task resets stayed fixed. The robot was not given a hardware descriptor.

Loading audited repeat results…

Compare repeat 1 with repeat 2 under the same declared setting. Watch whether the pastry is actually lifted and reaches the basket. A gripper reaching the basket is not enough. Playback follows the recording rate, not measured execution speed; videos end independently. All four raw views are available.

Wilson intervals are descriptive only. All eight arms use the same frozen modified RDP policy, four sampled plans, a 560-command limit, and action-plan updates every eight commands. Runtime 6ec00a56; audit f43d8b8f. All 64 recordings passed full decoding and hash checks. Initial-state inputs match; hidden simulator state and rendering are not bit-exact. Goal success does not establish safe success or tactile necessity.

Decision: do not train a timing network on these unstable preferences or expand the unchanged grid. Diagnose grasp retention before registering the next mechanism. This negative development result is not a general verdict against hardware conditioning.

Fewer visual updates can also lose a success.

The comparison below used a different reset and one attempt per setting. It is retained as context; do not pool it with the repeated hardware study above.

Does vision need to run on every command?

Both runs use the same frozen robot policy. One computes visual features from the two cameras on every command; the other reuses those features for eight commands. Both continue to receive fresh touch and robot state. Cameras still capture images.

Same starting state. Different visual timing.

Videos start together at their native recording rate; this is not a measurement of real-time execution speed. Each recording stops at its own endpoint. The restored renderer is used directly; no video denoising or generated frames were applied. Switch views to inspect the wrist camera or either tactile stream.

Fewer visual updates, one fewer observed success.

Wilson intervals are descriptive reference intervals. Two cases cannot establish equivalent population success rates. The successful episode stops early; failed episodes exhaust the command budget. Thus command totals here are not a like-for-like efficiency comparison among successful runs. An action command is one policy-client command, not a measured distance, energy cost, or physics substep.

Let local feedback decide when to update.

Inspired by Rodney Brooks, the proposed system combines a slow task planner with local visual and fingertip models. A learned controller would use their observations and the robot’s hardware properties to choose update intervals.

Three policies, the same starting states.

Six renderer-corrected episodes compare DP, modified RDP, and the decentralized hierarchy on reset 10. These methods differ in gates, primitives, and plan sampling. Their scores are context, not an isolated architecture comparison or a comparison against reset 11 above.