Differentiable Physics-Based System Identification for Robotic Manipulation of Elastoplastic Materials
Infers interpretable material properties from robot interaction and uses the identified model for downstream manipulation decisions.
Research direction
Experience-Efficient Physical Learning
Robots that learn useful physical knowledge from limited real-world experience — and know when another interaction is needed.
Real-world learningPhysical identificationInformative interactionDifferentiable physicsPhysical interpretability
Research focus
Real robots cannot assume unlimited, independently sampled training data. Physical interactions take time, wear hardware, can fail, and often reveal only part of what matters. I study how robots can learn continuously from the small amount of physical experience they can realistically acquire and turn that experience into knowledge that transfers beyond a single trial.
A central theme is active physical inference: identifying properties such as friction, compliance, geometry or latent material behaviour; recognising when current observations do not distinguish between plausible explanations; and selecting an interaction that resolves the uncertainty most relevant to action. My work combines learning with differentiable physics and system identification for deformable, granular and contact-rich manipulation.
I also use 3D and multimodal perception as enabling evidence sources. Geometry, motion, contact and material behaviour are not equally observable through every modality, so part of the problem is deciding whether the available observations are sufficient for the physical decision the robot needs to make.
I also treat interpretability as an operational property. A useful internal representation should connect a robot's belief to physical causes, uncertainty and testable interventions — not only provide a post-hoc verbal explanation after a decision has been made.
Current research
- Continual learning from small, sequential real-world datasets rather than assuming arbitrarily scalable robot data.
- Information-seeking interaction: deciding whether current observations are sufficient and what physical action would most usefully disambiguate competing hypotheses.
- Physically grounded and testable representations that connect learned knowledge to material properties, contact conditions, action consequences and uncertainty.
- Physics-aware learning and optimisation for contact-rich manipulation of deformable and granular materials.
- Reproducible and numerically reliable differentiable simulation for system identification, skill optimisation and physical design.
Selected published work
Selected work tracing the programme from physically informed learning to real-world identification and reliable differentiable optimisation.
DDBot: Differentiable Physics-Based Digging Robot for Unknown Granular Materials
Closes the loop from physical interaction and system identification to skill optimisation and real-robot execution in unknown granular media.
A Physics-Informed Demonstration-Guided Learning Framework for Granular Material Manipulation
Combines demonstrations, physical structure and learning to reduce the burden of acquiring manipulation experience.
Celebi's Choice: Learning Robotic Manipulation of Granular Materials through Differentiable Physics
Uses differentiable physical structure to learn and optimise granular-material manipulation.
On the Numerical Reliability of Differentiable Physics-Based Optimization for Robotic Material Manipulation
Studies gradient reproducibility, finite-difference validation and objective construction when physics is used inside robot inference and optimisation loops.