Joe Watson

I am a researcher in robot and machine learning. I am currently a postdoctoral research assistant at the Applied Artificial Intelligence Lab, Oxford Robotics Institute at the University of Oxford, working on world models and robot learning methods for sensorimotor manipulation. I am also affiliate with Kellogg College.

I received my PhD from the Technical Univesity of Darmstadt in 2024, working in the Intelligent Autonomous Systems group supervised by Prof. Jan Peters.

I studied Information Engineering at Peterhouse, University of Cambridge , where I was awarded the Charles Babbage senior scholarship.

Prior to starting my PhD, I worked on developing Versius, a novel robotic system for laparoscopic surgery, from prototype to product.

In 2022-23, I did an internship at Google DeepMind with the Robotics team, hosted by Sandy Huang and Nicholas Heess.

I am broadly interested in designing efficient inference-based algorithms for data-driven robotics using Bayesian inference and inductive biases.

Thesis

Research Highlights / Recent Research

Coherent Soft Imitation Learning
Joe Watson, Sandy H. Huang, Nicolas Heess (2023)
Advances in Neural Information Processing Systems (NeurIPS) [spotlight]
preprint / webpage / code / slides
TL;DR: A radical IRL algorithms that provable improves upon BC policies.

Inferring Smooth Control: Monte Carlo Posterior Policy Iteration with GPs
Joe Watson, Jan Peters (2022)
Conference on Robot Learning [oral].
paper / code / webpage / slides
TL;DR: Sample-based MPC with smooth priors and principled temperatures.

Differentiable Physics Models for Real-world Offline MBRL
Michael Lutter, Johannes Silberbauer, Joe Watson, Jan Peters, (2020)
IEEE Conference on Robotics and Automation (ICRA).
paper / code / webpage / slides
TL;DR: Recursive Newton Euler for differentiable SysID for RL.

Disentangling Dynamical Systems: Causal Representation Learning Meets Local Sparse Attention
Markus W. Baumgartner, Anson Lei, Joe Watson, Ingmar Posner (2026)
Conference on Causal Learning and Reasoning (CLEAR) [oral]
openreview /
TL;DR: Regularized in-context learning for model learning.

XQC: Well-conditioned Optimization Accelerates Deep Reinforcement Learning
Daniel Palenciek, Florian Vogt, Joe Watson, Ingmar Posner, Jan Peters (2026)
International Conference on Learning Representations (ICLR)
openreview / preprint / webpage / code
TL;DR: Batch norm, weight norm and a categorical critic is sample-efficient!

Global Tensor Motion Planning
An Thai Le, Kay Pompletski, Joao Carvalho, Joe Watson, Jan Peters (2025, 2026)
IEEE Robotics and Automation Letters, IEEE Conference on Robotics and Automation
preprint / code
TL;DR: Plan in parallel effectively.

Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization
Daniel Palenciek, Florian Vogt, Joe Watson, Jan Peters (2025)
Advances in Neural Information Processing Systems (NeurIPS)
openreview / preprint
TL;DR: Batch norm and weight norm is sample-efficient for deep RL!