James Heald

About Me

I develop algorithms for learning dexterous manipulation in high-degree-of-freedom hands, combining reinforcement learning and probabilistic machine learning. Most recently, this has led to an ICML 2026 Spotlight paper (top 2.2% of submissions), a NeurIPS 2026 paper, and a NeurIPS 2024 competition-winning method.

I am a Research Fellow at the Gatsby Unit, UCL working with Maneesh Sahani. I previously worked as a postdoctoral researcher with Daniel Wolpert at the Zuckerman Institute, Columbia University, where I published a first-author paper in Nature. I received a PhD in Engineering from the University of Cambridge, training in the Computational and Biological Learning Lab. Prior to my PhD, I worked as a medical doctor in the NHS, and before that, I received an MSc in Cognitive and Computational Neuroscience from The University of Sheffield and an MBChB in Medicine from The University of Manchester.

Open-Source Contributions

I am an active contributor to MyoSuite, participating in biweekly development meetings. Working collaboratively and independently, I have:

Ported environments from MuJoCo to MuJoCo XLA, enabling massively parallel, GPU-accelerated simulations.
Built an API to allow users to easily modify existing MuJoCo models for their own research needs.
Designed new models and tasks to expand the toolkit available to the wider research community.
Created an international advocacy website to widen participation in the NeurIPS 2024 MyoChallenge competition.

I have also made contributions to SBX, MuJoCo XLA and Brax.

Competitions

2024

MyoChallenge 2024: A New Benchmark for Physiological Dexterity and Agility in Bionic Humans

James Heald, Kai Biegun, Samo Hromadka, Maneesh Sahani

First Place ($2,000 prize)

Competition Paper

I lead a team that came first place in the Manipulation Track of the NeurIPS 2024 MyoChallenge competition. The challenge involved coordinating a biologically-realistic human arm with 63 muscles and a robotic arm with 17 DoFs to transfer objects between pillars. To control the musculoskeletal arm, we developed a novel algorithm to learn low-dimensional control manifolds in high-dimensional action spaces. This allowed us to efficiently acquire dexterous reaching and grasping skills. To control the robotic prosthetic arm, we combined a learned policy with inverse kinematic control, enabling smooth hand-offs and object relocation. Curriculum learning and reward shaping were used to solve the task in stages (reach, grasp, handover, move object to goal).

Descriptive Text for GIF — Our winning solution showcasing robustness to random pillar locations and object properties (weight, size and friction).

Descriptive Text for PNG — Our solution ranked first on all metrics: score, time, effort and peak contact force.

Select Publications

2026

Dexterous Skill Discovery via Topology-Aware Wasserstein Dependency

James Heald*, Kai Biegun*, Alexandre Galashov, Maneesh Sahani

Neural Information Processing Systems (NeurIPS), 2026

Poster Page Project Page Preprint Available on Request

We develop an unsupervised skill discovery method for learning dexterous skills by optimizing a novel Wasserstein dependency measure between skills and object motions. By formulating object motions in the Lie algebra 𝔰𝔢(3), the learned skills respect the non-Euclidean topology of object configuration space, enabling continuous object translation and rotation.

2026

Joint-Space Empowerment for Dexterous Coordination in Tendon-Driven Hands

James Heald, Vittorio Caggiano, Vikash Kumar, Maneesh Sahani

International Conference on Machine Learning (ICML), 2026

Spotlight (Top 2.2% of Submissions)

Also presented at 4th Workshop on Dexterous Manipulation: Scalable Learning for Human-Level Skills, Robotics: Science and Systems (RSS) 2026

Project Page ICML Paper RSS Workshop Paper Code Model Weights

We develop an information-theoretic framework for discovering low-dimensional action manifolds in tendon-driven systems. Across both robotic and musculoskeletal hand models, policies trained on these manifolds achieve substantially greater dexterity, sample efficiency and generalization.

2026

Maximum-Likelihood Learning of Latent Dynamics Without Reconstruction

Samo Hromadka, Kai Biegun, Lior Fox, James Heald, Maneesh Sahani

International Conference on Machine Learning (ICML), 2026

ICML Paper Project Page Code

We introduce the recognition-parametrized Gaussian state space model (RP-GSSM), which learns Markovian Gaussian latents via maximum likelihood without a decoder mapping latents back to observations. This lets the RP-GSSM learn task-relevant latent dynamics from video without auxiliary losses, outperforming alternatives even in the presence of background distractors.

2023

Contextual inference in learning and memory

James Heald, Máté Lengyel*, Daniel Wolpert*

Trends in Cognitive Sciences, 2023

Issue Cover and Feature Review

Paper Issue Cover

We review a computational framework for repertoire learning that provides a unifying account of phenomena across numerous domains, including motor learning, economic decision making, episodic memory and conditioning.

2021

Contextual inference underlies the learning of sensorimotor repertoires

James Heald, Máté Lengyel*, Daniel Wolpert*

Nature, 2021

Accompanied by Nature News and Views

Paper Code Data New and Views

We develop a theory of motor learning based on the principle of contextual inference. Our theory reveals that adaptation can arise by both creating and updating memories and changing how existing memories are differentially expressed.

Hobbies

When I'm not developing control algorithms, I enjoy executing biological ones: I love to learn calisthenics skills and rock climb (deep-water soloing, bouldering).