GUIDER on Real-Robot Manipulation Data

Intent prediction becomes useful when it identifies a plausible target early enough to support a person. This study evaluates the manipulation phase of GUIDER on recorded real-robot data, extending...

Intent prediction becomes useful when it identifies a plausible target early enough to support a person. This study evaluates the manipulation phase of GUIDER on recorded real-robot data, extending the earlier simulation work without claiming an end-to-end deployment of the complete mobile-navigation framework.

Why Real Recordings Add Evidence

Real sensing and arm motion contain characteristics that are difficult to reproduce perfectly in simulation. Evaluating recorded streams can reveal how an estimator behaves with that physical data while allowing the same evidence to be replayed consistently.

The distinction from closed-loop assistance is important. In replay, the estimator observes a trajectory that has already happened. An assistive controller would change the movement and potentially change the operator’s subsequent behaviour. The study evaluates the former, not the entire human response to the latter.

Retaining recorded timing is essential: using future frames or accelerating parts of a sequence could make an inference appear to arrive earlier than it did.

Refining Manipulation Candidates

The work adds online probability updates, workspace constraints, support-plane filtering and grasp-region prioritisation. These operations make candidate generation more closely reflect where a grasp could be useful.

Support-plane filtering helps separate a supporting surface from objects resting on it. Workspace constraints restrict the interpretation to relevant geometry, while grasp-region prioritisation moves beyond identifying an object to identifying useful parts of it. A visually prominent region is not necessarily an appropriate contact region.

Original paper view of manipulation-intent estimation during a medicine-grasping scenario
Original medicine-grasping visualisation from the real-robot GUIDER evaluation, Kenny, Contreras and colleagues, 2026. Unmodified, under CC BY 4.0.

Motion evidence then updates the relative plausibility of candidates. The output can be interpreted against a set of acceptable grasp candidates rather than requiring one unique pose to be correct. That distinction is fundamental when an object can reasonably be approached in more than one way.

Reported Evaluation

The study evaluates 20 manipulation steps across three assistance scenarios, including making tea and fetching medicine. Predictions remained within the correct grasp-candidate set in every evaluated case.

Reported quantity Value What it describes
Time to confident prediction 3.7 s When a confident intent estimate becomes available
Remaining time before first grasp 49.6 s Advance warning after confident prediction
Prediction stability 96.4% Persistence of the correct interpretation
Mean perceptual-phase runtime 4.857 s Processing cost of the perceptual phase
Median perceptual-phase runtime 4.474 s Middle runtime observation for that phase

The timing measures have different reference points. Advance warning is not algorithm execution time, and the perceptual-phase runtime should not be presented as the complete task duration or the rate of every probability update.

Correct membership in every evaluated candidate set is encouraging, but it is not a claim of perfect unique-grasp prediction for arbitrary users and objects. The evaluation set and its candidate definitions determine what that result means.

Interpretation and Remaining Questions

The results support the feasibility of obtaining early, stable manipulation-intent estimates from the recorded real-robot streams. They complement the earlier simulation study by changing the evidence source, rather than automatically validating every part of the earlier navigation-and-manipulation system.

A next question is whether acting on these estimates helps people complete tasks. That requires a closed-loop comparison in which assistance can affect the trajectory, with attention to incorrect interventions and recovery as well as successful ones.

Exact numerical reproduction needs the recorded sensor streams, transforms, task annotations, valid candidate sets, runtime environment and implementation. No such dataset is bundled with this explanation, and the reported averages should not be reconstructed from invented timelines.

Testing and Real-World Use

This could be tested by replaying time-stamped manipulation recordings with predeclared valid grasp sets, checking correctness, stability and the time remaining before contact without using future observations. In everyday assistive manipulation, early intent estimates could help a robot prepare relevant support for tasks such as fetching objects, subject to a separate evaluation of actual assistance.

Paper

The full preprint with robot photographs, scenario examples and timing figures documents the evaluation and its scope. Its results concern manipulation-phase inference on the recorded set.