Resource-Aware Space Robot Autonomy

The body changes the decision

A space robot can plan a collision-free motion and still be unable to carry it out. A joint may be too cold, a battery may not supply the peak current, or attitude control may have too little remaining authority. A more accurate perception model may also increase execution time and processor heat. These connections motivate the review's resource-aware view of embodied autonomy.

Embodiment concerns how sensing, structure, and dynamics shape behavior. Autonomy concerns which decisions can be made without immediate human direction. Learning estimates models or policies from data. These properties overlap without being interchangeable: conventional control can use rich physical feedback, and a trained neural controller can keep fixed parameters for an entire mission.

Make the resource contract explicit

Task planning proposes activities and priorities. Resource management returns feasible load modes and timing. Motion control supplies trajectories and enforces fast attitude, proximity, and contact limits. Their shared interface must carry usable constraints, not just isolated subsystem measurements.

Stored energy, allowable discharge power, component temperatures, and reaction-wheel momentum all constrain a candidate activity. Battery temperature and health affect the energy that can actually be used. A reserve must account for expected battery conditions and the energy needed for a safe transition.

Timing belongs in this contract. Fast motion and protection loops operate independently of slower forecasts. If scheduling misses a deadline, command arbitration needs an approved plan or a defined safe transition. Recovery must begin while sufficient control authority and resources remain.

Lower power is only half the calculation

Energy is power integrated over time. A processor with lower power draw can consume more energy per inference if it runs sufficiently slowly. Meaningful comparisons also include preprocessing, memory, idle periods, and any extra attempts caused by a reduction in model accuracy.

The spacecraft cares about completed work within constraints: whether a perception result arrives before the next control decision, whether peak current is available, and whether sustained operation stays within temperature limits. Processor power alone does not answer these questions.

Forecasting has a similar dependency on context. In the paper's illustrative margin example, a predicted peak of 79 °C against an 80 °C limit leaves little confidence if errors are typically 2 K. An average error is not a safety bound. The planner needs uncertainty characterized for its operating conditions, or a conservative alternative when those conditions are unfamiliar.

Three choices before eclipse

Consider the review's illustrative servicer preparing to capture a non-cooperative client before eclipse. The battery must support approach, contact, and recovery; the processor has limited heat rejection. Three options are available: capture now at full workload, capture now while deferring optional processing and communications, or wait for another opportunity. Navigation and collision avoidance continue in all three.

The full-workload option may retain enough energy yet exceed the processor's thermal limit after forecast error is included. The reduced-workload option can use the same geometric trajectory while reducing heat. If it also meets battery-power, reserve, and wheel-momentum limits, it can preserve the current capture window.

Waiting buys time to cool or recharge, but also incurs station-keeping and waiting loads. It must be evaluated from the state the system will reach after waiting. During execution, measurements update forecasts; optional loads can stop first, followed by a feasible hold or retreat if margins continue to shrink. This is a design example, not a reported flight outcome.

A supervisor must match the process timescale

A higher-level reasoning system can influence a slower physical process without issuing every actuator command. But the information window and response delay must match that process. The review describes ASTREA computing-load experiments in which supervisory analysis windows produced different thermal-control outcomes; the shorter window performed worse than the RL-only baseline in the reported comparison, while an orbit-aligned longer window improved the reported measures.

The lesson concerns supervisory timing within a bounded experiment. It does not establish that a particular model or update frequency is universally better. A thermal process, a contact controller, and a mission scheduler have different information needs and deadlines.

The next useful comparisons should measure which decisions a forecast changes, how much complete-task energy an implementation consumes, and whether recovery remains feasible. This connects algorithm performance to additional observations, maneuvers, or operating time on a real platform.