Project Showroom

DDBot combines a GPU-accelerated differentiable granular simulator with gradient-based optimisation. It identifies unknown material properties and optimises digging strategy without prior material knowledge, then transfers zero-shot to real robots.

Key contributions

• First-order gradient optimisation for differentiable granular simulation
• Efficient convergence (5-20 min) for both system identification and skill optimisation
• High-precision zero-shot deployment on sand and soil
• Parallel simulation with automatic differentiation

Preprint · Published Paper · Video Demo

DPSI estimates physically interpretable material parameters from one real robot interaction and incomplete point clouds, enabling accurate long-horizon simulation and manipulation planning for elastoplastic materials.

Key contributions

• Material parameter estimation from a single real-world interaction
• Works with incomplete and occluded 3D point cloud observations
• Interpretable physics parameters (e.g., modulus, friction, yield stress)
• High-fidelity simulation-to-real alignment for long-horizon planning
• Fully open-sourced workflow

Preprint · Published Paper · Video Demo · Project Page and Code

AutomaChef uses differentiable simulation to generate expert demonstrations through optimisation, then uses those demonstrations to train manipulation policies for granular transport with real-world transfer.

Key contributions

• Differentiable granular simulator implemented with Taichi
• Demonstrations generated automatically via gradient optimisation
• Outperforms standard RL, imitation learning, and prior task-specific methods
• Successful real-world granular transport deployment

Preprint · Published Paper

Celebi integrates causal inference into differentiable optimisation for granular manipulation, adaptively adjusting gradient step sizes to improve optimisation stability and convergence speed.

Key contributions

• Causality-enhanced gradient optimisation for granular skill learning
• Differentiable granular interaction simulation environment
• Differentiable skill-to-action mapping for trajectory optimisation
• Adaptive step-size control improves stability and convergence

Published Paper · Video Demo

GAM learns where and how to act in contact-rich disentangling tasks, combining affordance prediction with goal-conditioned reinforcement learning to handle diverse object configurations.

Key contributions

• General affordance prediction without task-specific engineering
• Generalises to diverse and unseen disentangling configurations
• Combines affordance learning and goal-conditioned RL
• Validated in real and simulated contact-rich scenarios

Preprint · Video Demo

A2E improves sparse-reward long-horizon robotic learning by combining abstract sub-task demonstrations with adaptive exploration that changes as each sub-task is mastered.

Key contributions

• Task decomposition into ordered sub-task sequences with abstract demonstrations
• Adaptive exploration strategy across mastered and weak sub-tasks
• Validated on grid-world and real robotic manipulation benchmarks
• Improves both sample efficiency and training stability

This work introduces Predictive Neighbouring Space Scoring to generate compact sub-goals in hierarchical RL for mapless navigation, reducing complexity compared with end-to-end learning from raw observations.

Key contributions

• Predictive Neighbouring Space Scoring (PNSS) for compact sub-goal generation
• Hierarchical RL with high-level goal selection and low-level control
• Lower learning complexity than end-to-end raw-sensor approaches
• Validated on challenging mapless navigation tasks

Published Paper · Video Demo

A survey of deep robotic affordance learning from an RL perspective, including taxonomy, limitations, and future directions focused on action-consequence prediction.

Key contributions

• Taxonomy of deep robotic affordance learning from RL perspective
• Connects affordance theory and modern deep RL
• Identifies open challenges in representation and deployment
• Proposes action-consequence prediction as a future direction

Preprint · Published Paper

The Universal Option Framework (UOF) decomposes long-horizon tasks into reusable sub-goals with universal low-level policies for efficient sparse-reward manipulation learning.

Key contributions

• Hierarchical decomposition of long-horizon tasks into reusable sub-goals
• Universal low-level policies generalise across sub-task instances
• Efficient learning under sparse rewards
• Demonstrated on challenging multi-step manipulation tasks

Published Paper · Video Demo · GitHub

An open-source PyBullet re-implementation of robotic multi-goal benchmarks with extra APIs and long-horizon tasks, removing dependency on commercial simulators.

Key contributions

• Open-source alternative to MuJoCo-based benchmark environments
• Added APIs for joint control, image observations, and cameras
• Introduced long-horizon sparse-reward benchmark tasks
• Validated parity against original benchmark behavior

Preprint · GitHub