The Oxford Robotics Institute (ORI) is built from groups of researchers, engineers and students all driven to change what robots can do for us. Industrial collaboration lies at the heart of their research agenda and ORI membership is the vehicle they use to achieve this, accelerating knowledge transfer from the ORI to its industrial members, currently including BP, Oxa, Accenture, Honda, L3 Harris, Navtech Radar and Scan Computers.

The ORI’s current research interests are diverse, from flying to grasping - inspection to running - haptics to driving, and exploring to planning. This spectrum of interests leads to researching a broad span of technical topics, including machine learning and AI, computer vision, fabrication, multispectral sensing, perception and systems engineering.

The ORI team named this two-phase model COMET (COmpetitive Mechanisms for Efficient Transfer) and it was evaluated both quantitatively and qualitatively, within image-based environments. Firstly, COMET was evaluated against two competitive baselines - Contrastive learning of Structured World Models (C-SWM) [1]; and Neural Production Systems (NPS) [3]. C-SWM learns a world model from observation via contrastive learning with a graph neural network (GNN) based dynamics model, whilst NPS learns to capture object interactions as independent mechanisms.

To perform these evaluations three distinct datasets were used - Particle Interactions, Traffic, and Team Sports. For each of these datasets, they defined a set of environments where objects can exhibit different behaviours. These environments were designed to test whether COMET could extract meaningful mechanisms and adapt to unseen environments via composition. The Particle Interactions dataset consisted of coloured particles that can interact with each other in different ways such as attraction and repulsion - defined by a combination of rules such as "red particles repel each other". The Traffic dataset contained observation sequences of traffic scenarios defined by traffic rules that apply to different vehicles such as "blue cars do not need to stop at red lights". The Team Sports dataset consisted of a simulated generic hockey game where players can perform different actions such as moving towards the puck or dribbling the puck towards the opponent goal, where some players might tend to take more aggressive actions. The below diagram shows how COMET was able to learn disentangled mechanisms that correspond to ground-truth behaviours in all three datasets, as indicated by the fact that each interaction mode has one main corresponding learnt mechanism. In contrast, NPS did not exhibit the same structure.

It was them tested as to whether COMET could apply these learnt mechanisms in new unseen environments. The below diagram illustrates using coloured tabs to show the ‘winning’ mechanism, and it can be seen that across all environments, the competition ‘winner’ changes as the underlying interaction mode changes. In the first row the particles repel each other when they are close (blue) and move independently when they are apart (green).

In the middle row the orange car obeys a slower speed limit and always picks the slow mechanism (orange), whereas the blue car approaches the red light with normal driving (pink) leading to slowing down (orange) and coming to a stop (green). Finally, in the bottom row the player first waiting to receive the ball (pink) and then moves towards opponent goal when in possession of the ball (orange).

Conclusion & Next Steps

The ORI team proved COMET was able to learn discrete, abstract concepts from diverse observations, and re-use such concepts to predict the evolution of unseen environments, and furthermore that COMET’s ability to perceive the world through the lens of structured high-level abstract concepts served as an important step towards efficient generalisation and transfer across different task settings.

Currently COMET lacks the ability to update the mechanisms during composition, which limits the model’s ability to adapt to environments with completely new interactions. Designing mechanism-based models that can initiate new mechanisms as they encounter new data was beyond the scope of this investigation, however the ORI team believe the method presented here will be instrumental in the development of such systems.

Further information on these experiments and their results can found by reading the full research paper.