Back to Research Robot Learning

Learning on Physically Adaptive Systems

Many manipulation tasks are too complex to hand-code: the behaviour is implicit, the state space too high-dimensional, or the environment too variable for classical controllers. Learning-based methods are one of the solutions — but they need data, and good data requires good control. Once a teleoperation platform is in place, it becomes a data collection engine. Human demonstrations captured through teleoperation feed imitation learning pipelines, and the resulting policies can execute tasks autonomously — closing the loop from human intent to robot behaviour.

1
From Demonstration to Autonomous Policy

A wide range of manipulation tasks — harvesting, inspection, leaf removal, monitoring, trimming, seeding — can be learned from human demonstrations and executed autonomously by a policy model. Compared to traditional visual servoing approaches, learning-based policies carry more implicit decision-making capacity: they generalise across object appearance, position variation, and scene complexity without explicit programming.

Harvesting
Inspection
Leaf Removal
Monitoring
Trimming
Seeding
vs. traditional visual servoing
Traditional: YOLO-based visual servoing

Detects objects and moves to them — but decisions are explicit and brittle to distribution shift.

Learning-based: autonomous object clearing

The policy autonomously clears every object it has learned — no explicit object list or sequencing rules needed.

Robustness to environment changes

Under an architecture designed to minimise sensitivity to environment variation, the policy remains stable under severe lighting disturbances.

HH Cheng, C Pan, A Desogus, Q Guan, L Sutter, J Hughes — Vision-Guided Agricultural Tasks with Diffusion-Based Motion Learning Using a Human-Robot Collaborative System. IEEE Transactions on Automation Science and Engineering (T-ASE), under review, 2025
2
Online Correction: Learning with Minimal Human Intervention

No policy is perfect from the start. When a robot fails — because the environment shifted, the training data was insufficient, or the task varied — the natural response is to dump more data and retrain. This work proposes a more efficient path: a human observes the failure, provides the minimal corrective input needed, and the robot continues autonomously. The result is an online learning loop that recovers from failures far more efficiently than passive data accumulation.

A human provides minimal corrective guidance the moment the robot reaches its limit.

After targeted correction, success rate exceeds what random data collection achieves with far more data.

C Pan, HH Cheng, J Hughes — Online Imitation Learning for Manipulation via Decaying Relative Correction through Teleoperation. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025
3
Human–Robot Collaboration

Autonomy and human control are not opposites — they exist on a spectrum. The most capable systems know when to act alone and when to defer. This demonstrates a robot that operates autonomously while remaining aware of human presence and intent, adjusting its behaviour to collaborate naturally in a shared workspace.

The right arm is human-controlled via teleoperation; the left arm runs the same autonomous policy as above — both operating simultaneously in a shared workspace.

Where This Is Going

Three directions building on this foundation:

Robot-Assisted Teleoperation
Instead of the human controlling everything, the robot understands the task and scene — and predicts what the operator is trying to do. It offers assistive corrections in real time, reducing cognitive load and improving precision. The human stays in control, but the robot becomes a co-pilot rather than a passive tool.
Learning the Morphing Body
Rather than learning to complete a specific task, the robot learns what its own body is capable of — and how to reconfigure it to best suit the task at hand. This is particularly relevant for reconfigurable cable robots, where the physical structure itself is a design variable. The policy learns the body, not just the behaviour.
Foundation Models for Large-Scale Manipulation
Cable robots and multi-arm systems operate over workspaces too large and too varied for task-specific policies. The goal is a foundation model that understands spatial task context at scale — one that can reason about where to go, what to grasp, and how to coordinate across a large workspace, enabling zero-shot or few-shot adaptation to new environments and tasks without retraining from scratch. Under preparation.