Research

How machines come to act.

Our work is organised around the perceive–plan–act loop. We treat the loop as a single system rather than three separate modules, and ask what changes when each stage is designed to be uncertain, reactive, and transferable.

01 · Perception

Perception

Robots need more than labels — they need representations they can reason over. Our work on perception centres on scene understanding that carries calibrated uncertainty, so downstream planning can distinguish what the robot knows from what it only suspects.

Object- and relation-centric scene representations for manipulation and navigation.
Predictive uncertainty for perception under occlusion, clutter, and changing light.
Multi-modal fusion of vision, depth, and proprioception into a shared belief state.

Representative work: Perceive–Plan–Act with Predictive Uncertainty in Cluttered Scenes (ICRA 2025).

02 · Planning

Planning

Planning is the bridge between what a robot believes and what it does. We study both long-horizon task planning and real-time motion planning, with an emphasis on methods that remain tractable when the robot’s belief state is uncertain.

Sampling-based and learned planners for cluttered, dynamic environments.
Planning directly over probabilistic representations rather than point estimates.
Task-and-motion planning that composes high-level goals with low-level constraints.

Representative work: Neural Potential Fields for Real-Time Motion Planning at Scale (CoRL 2024).

03 · Control

Control

Control turns plans into physical action. Our interest is in learned, whole-body policies that keep the stability guarantees classical control provides, so that a policy trained for one task does not collapse the moment the world pushes back.

Reactive whole-body control for legged and mobile manipulators.
Learning policies from demonstration, with residual classical structure where it helps.
Safety and stability constraints carried through from training to deployment.

Representative work: Learning Reactive Whole-Body Control from Language-Conditioned Demonstrations (RSS 2024).

04 · Sim-to-real

Sim-to-real

Simulation lets us train what hardware could never afford. The open question is transfer: how much of what a policy learns in simulation survives contact with a real robot. We work on the gap itself, treating it as a research problem rather than an inconvenience.

Domain-randomised and adversarial training that carries across the gap.
Latent world models that let a robot plan against imagined futures it has actually seen.
Benchmarks that measure transfer explicitly, instead of assuming it.

Representative work: Latent World Models for Contact-Rich Manipulation (CoRL 2025) and Closing the Sim-to-Real Gap via Adversarial Domain Distillation (IROS 2025).

Output

Where the work lands.

Every area is tied to published results and, where possible, to a released dataset or benchmark so the claims can be checked.

See the publications