Adaptive Robotic Manipulation Skills

Research Context and Potential Impact
Service robots have the potential to transform domains such as industrial maintenance and domestic care. To operate effectively in unstructured environments, they must generalize across diverse tasks based on environmental observations and minimal task descriptions.
While some methods generalize within a task class (e.g., opening drawers), handling fundamentally different tasks (e.g., pressing a button vs. tightening a valve vs. rearranging objects) remains a challenge. Robots must be able to interpret semantic task descriptions and synthesize appropriate manipulation strategies for novel tasks.
Further complexity arises from tasks requiring non-visual information, such as detecting when a valve is tight. These tasks demand that robots quickly adapt plans in response to unexpected visual and non-visual changes -- potentially altering strategies or tools mid-execution. Enabling this level of adaptability remains an open problem.
Aims and Objectives
The project will investigate algorithmic and learning-based methods for synthesizing policies that enable robots to generalize across manipulation tasks in real-world, unstructured environments. This includes synthesis from semantic, conceptual descriptions of the task written in natural or formal language. We plan to use reinforcement learning (RL) to synthesize these policies.
The project will also explore adaptive robot behaviors that respond to unexpected visual and non-visual changes by preempting failures and adjusting strategies based on data gathered through repeated task interactions.
The proposed methods will be extensively evaluated on real robots in the lab.
Novelty of the Research Methodology
Learning Generalizable Robot Policies for Long-Horizon Tasks
Linear Temporal Logic (LTL) has recently emerged as a powerful formalism for specifying complex, temporally extended tasks in multi-task RL. Recent work has shown that policies can be trained to generalize across LTL specifications, which in turn can be synthesized from natural language commands. However, these specifications rely on a fixed, discrete vocabulary of atomic propositions known a priori, limiting their applicability in real-world, unstructured environments. We will develop vocabulary representations that scale effectively to unstructured, real-world robotics environments.
In addition, prior approaches have been restricted to environments with simple dynamics (e.g., point agents) and access to ground-truth task information -- conditions that do not reflect realistic robotic scenarios. We will extend these methods by incorporating techniques from robotics for learning complex manipulation skills and perceptually grounding task progress.
Adaptive Robot Behavior
Several task and motion planning (TAMP) frameworks allow robots to adapt to dynamic environments by generating plans conditioned on the current scene. However, this adaptability is typically limited to visually observable changes (e.g., real-time collision avoidance). This project extends that scope to include changes that are not visually observable (e.g., a value becoming tight).
In addition, we will explore how a robot can adaptively select and switch between different manipulation strategies for the same task based on prior experience. Each strategy carries trade-offs (e.g., speed vs. max applied wrench), and different task instances may require different approaches (e.g., stiffer valves requiring more wrench). To our knowledge, this kind of strategy-level adaptation within task classes has not been addressed in prior work.
Alignment to EPSRC Strategies and Research Areas
This project directly relates to the Robotics and AI EPSRC research areas.
Involvement of Companies or Collaborators
The project will be supervised by Prof. Ioannis Havoutis at the Oxford Robotics Institute (ORI) and co-supervised by Prof. Alessandro Abate at the Oxford Control and Verification group (OXCAV).