Record operator haptic feedback as a demonstration-collection condition, separately from the observations supplied to a learned robot policy. Feedback can change how a person demonstrates an action; robot-side tactile input changes what the deployed model can observe. A comparison that changes both cannot isolate either contribution.
The HapTile paper, version 2, revised June 19, 2026 and checked October 15, describes fingertip tactile sensing on the robot and haptic feedback to the teleoperator. Its marker-motion signal drives discretized controller vibration during active teleoperation. The paper identifies limited feedback resolution and collection within a single laboratory scene as limitations. No data collection or policy experiment was performed here.
A tactile-policy ablation does not isolate operator feedback
The paper’s policy-modality comparison uses vision-only, vision plus tactile images, and vision plus tactile images and marker-motion features. Its vision-only baseline still receives robot proprioception alongside third-person and wrist RGB images; ‘vision-only’ means no tactile input in this comparison. Those conditions change policy observations. They should not be renamed as a controlled demonstration-collection comparison with haptic feedback enabled versus disabled.
For an illustrative collection study, the same operator could demonstrate a contact-rich task with and without the feedback treatment. Match the robot sensing and action logging, task instruction, intended initial-state distribution and demonstration budget. Record trial order and operator identity so practice or fatigue is not silently attributed to vibration. This is a proposed study, not a HapTile result.
Inspect the feedback mapping and thresholds as well as its enabled state. The documented signal conveys discretized contact information; it does not supply a continuous calibrated force measurement to the operator. Keep reset/setup intervals distinct, since the paper enables feedback only during active teleoperation.
Separate collection treatment from deployment observation
For a proposed downstream comparison, hold the policy inputs, training budget and evaluation protocol fixed while changing the demonstration-collection condition. A second comparison can hold the collected demonstrations fixed while changing the policy’s tactile inputs. If the available datasets cannot support both controls, state which question remains unresolved.
Retain the commanded action, tactile observations, left and right feedback values, enabled state and task outcome with each source episode. Do not infer that an autonomous rollout receives operator feedback just because that feedback helped produce its training data. Also keep failure and correction history: more confident demonstrations can still contain unsuitable actions or inconsistent task outcomes.
For deployment, evaluate the resulting policy without undisclosed operator assistance under a stated task and scene protocol. The single-scene collection limit matters when claiming robustness to different layouts, lighting or backgrounds. The teleoperation-data guide covers synchronized demonstration records; this HapTile-specific check adds what the demonstrator could feel and prevents that collection condition from being confused with the robot’s inference-time sensors.