Peer-reviewed and preprint work from 2024 onward, listed with venue and year. Where a paper has an accompanying release, it is noted alongside.
A benchmark that evaluates perception, planning, and control together rather than in isolation, measuring how errors in one stage propagate through the full loop. Released with the FlowBench suite in our open-source repository.
We study a single token vocabulary shared by all three stages of the loop, and what is gained — and lost — when perception, planning, and control read and write the same representation.
A latent world model trained to predict the effect of contact-heavy actions, letting a manipulator plan against imagined futures for insertion and assembly tasks where analytic models fall short.
We treat the sim-to-real gap as an adversarial problem, distilling a policy through a discriminator that is trained to tell simulated rollouts from recorded hardware runs, and report transfer without task-specific tuning.
An architecture that carries calibrated predictive uncertainty from perception through planning, so the robot can prefer trajectories whose outcomes it can actually foresee — and hesitate where it cannot.
Whole-body policies learned from language-conditioned demonstration, retaining reactive behaviour so the robot can recover from disturbance instead of replanning from scratch.
Learned potential fields that trade the global guarantees of classical planners for real-time behaviour in large, dynamic environments, with the failure modes we observed reported alongside the results.