ICML 2026 Demystifying Action Space - Heungwoo/research GitHub Wiki

Demystifying Action Space Design for Robotic Manipulation Policies — A 13k-rollout action-space study

Venue: ICML 2026 (Poster) Category: Analysis-Insight Affiliations: Tsinghua (IAIR/Wuxi) Traction (2026-06-09): 1 citation (arXiv preprint)

📖 In-depth review (taxonomy, EEF-vs-joint verdict, full findings, figures): Review-Demystifying-Action-Space

Problem

The action space fundamentally shapes the optimization landscape of imitation-based manipulation policy learning, yet while the field has poured effort into scaling data and model capacity, action-space choices remain guided by ad-hoc heuristics or legacy designs. This leaves an ambiguous understanding of what action representations are actually good and why.

Method

Large-scale systematic empirical study confirming that the action space has significant and complex impacts on policy learning.
Dissects the action design space along temporal and spatial axes, enabling structured analysis of learnability and control stability.
Compares absolute vs. delta action representations and joint-space vs. task-space parameterizations.
Grounds the analysis in real-robot data rather than simulation alone.

Results

The study is based on 13,000+ real-world rollouts on a bimanual robot and evaluation on 500+ trained models over four scenarios. Key findings: predicting delta actions consistently improves performance, while joint-space and task-space representations offer complementary strengths, joint-space favoring control stability and task-space favoring generalization. The abstract presents these as trade-off conclusions rather than a single headline accuracy number.

Significance

This is an insight/analysis paper, not a new architecture: it replaces ad-hoc action-space heuristics with empirically grounded design guidance from one of the largest real-robot ablation efforts to date. The delta-action recommendation and the joint-vs-task-space stability/generalization trade-off give practitioners concrete defaults. As 2026 VLA and imitation pipelines standardize, having the action representation choice demystified at this scale is a useful piece of foundational evidence.