Remote battery disassembly requires both an understanding of the workcell and fine control of physical interaction. This paper combines extended-reality visualisation, haptic feedback and learned motion guidance in a variable-autonomy framework, investigating how the robot can assist an operator without reducing the task to either fully manual or fully autonomous control.
Three Different Kinds of Assistance
Extended reality supplies spatial context: it helps the operator relate the remote robot, tool and battery components. Haptic feedback communicates through forces, supporting a different channel of perception and guidance. Learned motion models provide a task-related reference for movement.
These functions should not be conflated. A clearer visualisation does not itself generate a trajectory, and a useful trajectory does not tell an operator what contact feels like. Their combination is valuable because teleoperation involves seeing, deciding and acting through an indirect interface.
The framework is implemented on an industrial KUKA manipulator and evaluated in structured tasks and real EV battery-disassembly tasks. The experimental setting matters when interpreting its performance: it is not a claim that the same configuration can handle every battery design or condition.
Learning Motion Relative to the Task
Task-parameterized Gaussian mixture regression, or TP-GMR, represents demonstrated motion in relation to task frames. A frame describes a relevant spatial reference, such as an object or work region. Expressing motion relative to those references provides a way to adapt learned guidance when the task geometry changes.
Regression uses the learned distribution to produce an estimated motion. The purpose is not simply to replay one demonstration at the same absolute coordinates. Task-relative information lets the guidance reflect where relevant parts of the task are now, subject to the assumptions and demonstrations used to learn it.
The framework also prepares corrective trajectories. Precomputing these responses reduces the delay associated with obtaining a correction when needed. This is an architectural change to a planning operation, not a claim that physical manipulation can occur instantaneously.
graph TD
A[Demonstrations and task frames] --> B[TP-GMR motion guidance]
B --> C[Prepared corrective trajectories]
D[Operator input] --> E[Variable-autonomy control]
C --> E
F[Modelled spatial constraints] --> E
E --> G[Robot task execution]
G --> H[XR view and haptic feedback]
H --> D
Original explanatory diagram of the roles described in the paper, not the paper’s controller block diagram.
Constraints and Human Control
Constraint barrier functions enforce the spatial restrictions represented in the framework. That provides a defined boundary for the commanded behaviour, but only for hazards and dynamics captured by the model. It is not a blanket guarantee covering unmodelled contacts, electrical hazards or calibration failures.
Variable autonomy changes the balance between operator input and robot assistance. The important interaction question is whether the guidance helps the person perform the task while remaining understandable and controllable. A combined interface comparison can evaluate the whole experience, but cannot automatically assign each improvement to one component.
Reading the Performance Claims Correctly
The paper reports that precomputation reduced replanning response from 2.0 seconds to under one millisecond. This concerns the measured correction/replanning operation. It is not human reaction time, camera-to-motion latency or total task duration.
Relative to manual teleoperation without assistance, the reported experiments achieved task-time reductions of up to 48% and a 32% reduction in path deviation. “Up to” identifies the reported maximum improvement, not an expected reduction for every trial or operator.
These results motivate further study of combined spatial and force guidance. Exact reproduction would require the demonstrations, task frames, controller settings, calibrated interfaces, robot geometry and study protocol. Keeping planning response, trajectory quality and participant task performance as separate outcomes is essential to an informative comparison.
Testing and Real-World Use
This could be tested first with simulated manipulation trajectories, comparing unassisted and guided operation on matched tasks while measuring completion time, path deviation and intervention needs separately. In supervised remote disassembly, the approach could help an operator follow useful task motions and perceive interaction through visual and force cues, with physical battery work remaining subject to qualified laboratory procedures.
Paper and Original Figures
The Journal of Field Robotics article appeared online on 25 September 2025 and in the March 2026 issue. The Aston University record provides the abstract and an open-access full paper containing the original experimental figures.