A mobile manipulator has to travel through an environment and then interact with it. Asking a person to control every base and arm movement can be demanding, but handing everything to autonomy can fail when the environment or task becomes uncertain. This mini-review examines how variable autonomy distributes that responsibility between a human and a robot.
Variable Autonomy Is More Than an Autonomy Slider
The review synthesises 38 included papers. It is a map of approaches and open problems, not a new experiment establishing that one control policy wins. Its search and exclusion information describes how the reviewed evidence was selected, and its conclusions describe the literature available at publication.
A central distinction is between who can change autonomy and how control is shared. In human-initiative systems, the operator chooses an autonomy change. In AI-initiative systems, the machine initiates it. Mixed initiative permits both to contribute, but still requires rules for accepting changes and resolving conflicts.
Shared control is a different concept: human and machine contribute concurrently to the command. An operator might steer while the robot modifies motion near an obstacle without any discrete change of mode. Calling every assisted movement an autonomy switch would blur these different designs.
graph TD
A[Human-robot task] --> B[Who can initiate an autonomy change?]
B --> C[Human initiative]
B --> D[AI initiative]
B --> E[Mixed initiative]
A --> F[How is control contributed?]
F --> G[Manual, autonomous or concurrent shared control]
A --> H[Which subsystem?]
H --> I[Mobile base, arm or coordinated whole body]
Original explanatory diagram separating three design questions discussed by the review; it is not a reproduced paper figure or a ranking of systems.
Why the Base and Arm Complicate the Problem
Navigation and manipulation have different immediate objectives. The base must find a useful, reachable position, while the arm must carry out a local interaction. Their decisions remain coupled: a poor base position can make the target unreachable, and moving the base during contact can disturb the arm’s task.
A system can therefore have different autonomy levels for its base and arm at the same moment. Describing the whole robot with a single label can conceal who controls the component that matters in a particular phase. Whole-body coordination asks how those decisions should be made together rather than treating locomotion and manipulation as unrelated stages.
This is especially relevant in cluttered or hazardous settings, where simply repositioning the platform may require careful supervision. The review identifies integrated whole-body variable autonomy as a research direction, not an already solved general-purpose capability.
Workload and Communication
Assistance can reduce repetitive input, but it can also create a new monitoring burden. If the operator cannot understand a robot’s current mode or anticipated action, intervention may become harder precisely when it is most needed. Workload measures therefore belong alongside task success and completion time.
Communication conditions are another part of the task, not an incidental implementation detail. Delay, limited bandwidth and missing information can make direct teleoperation difficult. Local autonomy and predictive information may help, but the appropriate response depends on the task and the operator’s ability to recognise and override an unsuitable action.
These concerns interact. A system that changes modes frequently might improve one movement measure while making its behaviour harder to follow. Conversely, a predictable but less aggressive assistance policy may be easier to supervise. The review motivates evaluating these trade-offs explicitly rather than equating increased autonomy with better collaboration.
What Can and Cannot Be Concluded
Differences in environments, tasks, interfaces and evaluation measures limit direct comparisons between studies. A policy demonstrated during navigation in one setting is not automatically validated for contact manipulation elsewhere. Likewise, a reported workload reduction does not by itself establish improved safety.
The review is useful as a vocabulary and evidence-organising tool: it helps identify the task, the controlled subsystem, the authority to change modes and the outcomes actually measured. Repeating the literature review would require rerunning its documented search and screening process with a recorded search date; subsequent publications and indexing changes may produce a different included set.
Testing and Real-World Use
This could be tested by comparing clearly specified autonomy policies on the same mobile-manipulation task while recording success, workload, interventions and communication conditions. In remote inspection or maintenance, the review can help designers choose where assistance is valuable and make responsibility for base and arm control understandable to the operator.
Full Review
The open-access Frontiers review contains the literature-selection details and references needed to trace each reviewed approach to its original study.