Video Models for VLA • Scalable Robot Learning • Offline RL
PhD Candidate, RoMA Lab, UCL
I research foundational AI for robotics, focusing on video model backbones for Vision-Language-Action models. I believe video generation architectures encode dynamics in ways VLMs cannot, enabling better generalisation and scaling toward embodied AGI. Supervised by Chris Xiaoxuan Lu and Dimitrios Kanoulas. Funded by EPSRC Landscape Award.
arXiv 2024
Supervised learning approaches to offline reinforcement learning, particularly those utilizing the Decision Transformer, have shown effectiveness in continuous environments and for sparse rewards. However, they often struggle with long-horizon tasks due to the high compounding error of auto-regressive models. To overcome this limitation, we go beyond next-token prediction and introduce Planning Tokens, which contain high-level, long time-scale information about the agent's future. Predicting dual time-scale tokens at regular intervals enables our model to use these long-horizon Planning Tokens as a form of implicit planning to guide its low-level policy and reduce compounding error. This architectural modification significantly enhances performance on long-horizon tasks, establishing a new state-of-the-art in complex D4RL environments. Additionally, we demonstrate that Planning Tokens improve the interpretability of the model's policy through the interpretable plan visualisations and attention map.
Next-gen benchmark for dexterous bimanual manipulation. 11 long-horizon tasks on a compact table-top with mobile camera, distractor objects, and low-cost compliant arms. Designed to expose the gap between saturating sim benchmarks and real-world humanoid performance.
Turn your webcam into real-time robot joint positions. Designed for LeRobot with Wilor GPU backend, Kalman smoothing, and plug-and-play integration. 48 GitHub stars.
Multi-agent many-shot code generation achieving 4th place out of 800+ participants in the NeurIPS 2024 Meta Hackercup AI Track. Extended MapCoder with careful prompt engineering, Codestral-22B, Maj@128 voting, and VLLM parallel inference (2000 tk/s).
Self-supervised hint distillation for improving LLMs on code generation. Finetunes Qwen3-4B using KL-divergence distillation where the model with hint access teaches itself.
Prime verifiers environment for VPCT scenes. Won #3 in the Iterate London RL Environment Hackathon (Dec 2025). Scores model outputs predicting final bucket positions.
A $450 low-cost 7-DOF robot arm bridging the gap between ultra-budget ($140 SO101) and premium ($3000+) options. 1.2kg peak payload, 60cm reach. Designed for accessible robot learning research.
Contributor to HuggingFace's open-source VLA framework for robot learning. Contributed BlockPush environment for benchmarking manipulation policies, assisted in porting HIL-SERL for online RL finetuning, and improved training speed through data loading optimizations.
State-of-the-art latent diffusion model trained on FFHQ for photo-realistic face generation up to 128x128px. Optimized for single GPU training. 100% coursework grade.
Hackathon project detecting when housemates leave dirty dishes using a hybrid of YOLO and traditional computer vision state tracking algorithms.
Open source snap-on camera mount for SO100 robot arm and U20CAM camera. Optimized 30-degree angle based on testing. Parametric Fusion 360 design for easy customization.
AI-first qualitative coding platform for transcript analysis. Helping consultants efficiently analyze interview and research data.
Non-profit IB revision platform I founded and maintain. 1.6 million page views, #4 Google result for "IB revision". Full-stack PHP/SQL with custom CMS.
First complete 3D graphics and physics engine for Scratch. Built over 2 years with innovative binary space partitioning for efficiency. Demonstrates deep graphics understanding.