ICML 2026 Joint Space Empowerment as a Theory - Heungwoo/research GitHub Wiki

Joint-Space Empowerment as a Theory of Dexterous Motor Coordination — an information-theoretic principle for overactuated control

Venue: ICML 2026 (Poster) Category: Dexterous Affiliations: authors James Heald, Vittorio Caggiano, Vikash Kumar, Maneesh Sahani

Problem

Searching for effective policies in high-dimensional action spaces is notoriously challenging. This difficulty is compounded in overactuated musculoskeletal systems, where multiple muscles span each joint, and individual muscles actuate multiple joints. Although this redundancy complicates naive policy search, it also implies that effective control can be captured by a low-dimensional action manifold.

Method

To identify such a manifold, the authors introduce joint-space empowerment (JSE), a novel information-theoretic principle that quantifies how much control an agent has over its body. JSE is used to discover high-empowerment action manifolds — low-dimensional subspaces of the full muscle-actuation space on which the agent retains maximal influence over its joint configuration. Manipulation policies are then learned directly on these discovered manifolds rather than in the raw, redundant action space.

flowchart LR
    A[Overactuated musculoskeletal body] --> B[Compute joint-space empowerment]
    B --> C[Discover high-empowerment, low-dim action manifold]
    C --> D[Learn manipulation policy on manifold]
    D --> E[Enhanced dexterity, sample efficiency, generalization]

Results

Manipulation policies learned on the high-empowerment manifolds show significantly enhanced dexterity, sample efficiency, and improved generalization compared with learning in the full action space. The abstract reports these gains qualitatively rather than as headline numbers.

Significance

These results suggest a general principle for motor coordination in high-dimensional, overactuated systems, with implications for both biological motor control and embodied artificial agents. By grounding action-space reduction in an information-theoretic notion of control authority (empowerment) rather than hand-designed synergies, JSE offers a principled route to taming muscle-level redundancy — connecting computational motor neuroscience with practical dexterous policy learning.