GAM: General Affordance-Based Manipulation for Contact-Rich Object Disentangling Tasks
Moves from object labels toward reasoning about action possibilities and consequences in contact-rich manipulation.
Research direction
Self-Aware Resilient Robotics
Robots that can establish what body and capabilities they actually have, detect when these change, and reconfigure safely.
Embodiment discoveryCapability identificationDamage adaptationResilient autonomyBounded agents
Research focus
Most robot autonomy begins after engineers have already defined the embodiment: the joints, sensors, actuators, interfaces, limits and software stack are assumed to be known. I am interested in the layer before that assumption — how an autonomous system can establish an operational model of itselfsufficiently well to configure, validate and deploy useful behaviour.
Operational embodiment goes beyond morphology. It includes the robot's physical structure, sensing and actuation interfaces, reachable actions, capability envelope, uncertainty, operating limits and current health state. This creates a natural route from embodiment discovery to resilience: when a joint degrades, a sensor shifts or a tool changes, the robot should detect that its previous self-model is no longer adequate, identify what capabilities remain valid, and adapt or escalate safely.
Foundation models and agentic tools can be useful components, but I treat them as bounded engineering tools rather than the research end-point. The scientific problem is how to generate evidence that a newly configured or adapted robotic system is operationally valid before and during physical deployment.
Current research
This is a growing research programme. The publications below are earlier foundations in affordance/capability reasoning and reusable robot control; current work extends these ideas toward explicit embodiment discovery, health-aware capability estimation and recovery.
- Operational embodiment discovery: identifying a robot's body, sensors, actuators, interfaces, limits and usable action spaces from available evidence and targeted tests.
- Capability identification and validation: determining not only what hardware exists, but what tasks and actions are currently feasible with sufficient confidence.
- Embodiment damage and degradation adaptation: detecting inconsistency, diagnosing changed capability, re-identifying what remains possible and recovering conservatively.
- Bounded AI-assisted configuration, testing and redeployment of robot control systems with explicit verification gates and human authority.
- Action-consequence and affordance reasoning as a bridge between physical state, embodiment and usable robot capability.
Selected published work
Earlier work that provides the conceptual and technical foundations for capability-centred and adaptive autonomy.
Recent Advances of Deep Robotic Affordance Learning: A Reinforcement Learning Perspective
Connects affordance representations, robot learning and deployment, providing a foundation for capability-centred autonomy.
Hierarchical Reinforcement Learning with Universal Policies for Multi-Step Robotic Manipulation
Develops reusable hierarchical policies for long-horizon manipulation, an early step toward robot systems that can compose capabilities rather than relearn each task from scratch.