Review Demystifying Action Space

Demystifying Action Space Design — In-Depth Review (EEF vs Joint, Absolute vs Delta)

Venue: ICML 2026 · arXiv: 2602.23408 Affiliations: Tsinghua University (IAIR / Wuxi Research Institute of Applied Technologies) Scale: 13,000+ real-world rollouts · 500+ trained models · 14 tasks (4 real + 10 sim) · 3 platforms One-pager: ICML-2026-Demystifying-Action-Space · See also VLA Architectures

TL;DR

This is the large-scale empirical study that asks: for an imitation-learning manipulation policy, should the action be in joint space or end-effector (task) space, and should it be absolute or delta? The answer, from 13k+ real rollouts, is not "one wins" — it is a decision rule:

• Temporal axis (absolute vs delta): delta wins, almost always — but only if implemented chunk-wise, not step-wise.
• Spatial axis (joint vs EEF): complementary. Joint space → control stability and gets better with scale; task/EEF space → generalization (cross-embodiment, transfer).
• Practical rule: single platform, maximize performance → joint + chunk-wise delta. Cross-embodiment / transfer → task-space (EEF).

---

1. Why this study

Modern VLA / IL work pours effort into scaling data and model size, but the action space — the very thing the policy regresses to — is still chosen by "ad-hoc heuristics or legacy designs" (RT-1 used discretized EEF; π uses joint deltas; ACT uses joint absolute; etc.). The paper argues the action space "fundamentally shapes the optimization landscape," and sets out to measure it under controlled conditions instead of folklore.

2. The action-abstraction taxonomy (§2)

The design space is factored into two orthogonal axes plus an implementation detail:

Spatial abstraction — where the action lives

• Actuator space (torque / current) — the low-level controller layer.
• Joint space (configuration, ℝⁿ) — joint angles; reached from EEF via inverse kinematics (risk: singularities).
• Task space (end-effector SE(3) pose) — the gripper pose; reached from joints via forward kinematics.

Temporal abstraction — how the action is referenced in time

• Absolute (0th order) — predict the global target position.
• Delta (1st order) — predict the relative displacement from a reference (needs integration to execute).
• Higher-order (accel / force) — requires accurate inertial modeling and high-frequency feedback (out of scope for standard IL).

Action chunking — the decisive implementation nuance

When a policy predicts a chunk of future actions, a delta can be referenced two ways:

• Step-wise delta — each step relative to the previous predicted step (errors integrate along the chunk).
• Chunk-wise delta — every step relative to the single anchor frame at the chunk's start (errors stay independent).

3. Experimental setup (§3)

• Policies: ACT (regression backbone) and DP (flow-matching / diffusion policy) — so conclusions aren't tied to one model class.
• Platforms: single-arm AgileX, bimanual AgileX, and RoboTwin 2.0 simulation.
• Tasks/data: 14 tasks (4 real + 10 sim); 250 demos/task (real), 50/task (sim); trained to 600 epochs (and 900/1200 for scaling).
• Scale knobs: data {100, 250, 500} trajectories × compute {600, 900, 1200} epochs; single- and multi-task; transfer from π0; cross-embodiment.

4. Findings

RQ1 — Implementation nuances are decisive: chunk-wise ≫ step-wise delta

Chunk-wise delta consistently and significantly beats step-wise delta — the gap reaches upwards of 10% on average. The paper gives the mechanism as a stability theorem:

Proposition 4.1 (noise amplification). For a predicted chunk of length k with bounded per-step noise, the decoded execution error scales as O(k) for step-wise delta (the transform is the lower-triangular all-ones matrix Lₖ, so errors accumulate along the horizon), but stays O(1) for chunk-wise delta and absolute actions (the transform is the identity Iₖ, so errors propagate independently).

So step-wise delta structurally amplifies prediction noise as the horizon grows; chunk-wise does not. Takeaway: if you use delta actions, always anchor them chunk-wise.

Horizon coupling: absolute control prefers a longer execution horizon; delta control peaks at a shorter horizon (relative reps are sensitive to execution drift). They train at k=60 (2 s @ 30 Hz) and grid-search the execution window 15–60 — i.e., the chunk horizon is not a constant, it must be tuned to the temporal abstraction.

RQ2 — Systematic trends

Temporal (delta vs absolute): with the best implementation of each, delta still consistently and significantly outperforms absolute across all platforms, tasks, and both ACT and DP. Why: (1) mapping high-dim vision → global coordinates has low local coherence, whereas predicting immediate displacement is a more tractable inductive bias; (2) absolute reps need long horizons that are hard to train.

Spatial (joint vs EEF) — the headline comparison: the two are complementary, not ranked:

• Joint space → the more robust / stable spatial representation in most single-platform cases.
• Task space (EEF) → competitive in low-data/limited-compute regimes and better for generalization.

RQ3 — Consistency under scaling, transfer, cross-embodiment

• Delta stays superior as data and compute scale (sometimes by only a marginal margin).
• Joint-space superiority becomes more pronounced as epochs and data increase (especially for regression policies) — it "benefits disproportionately from stronger modeling and extensive training," apparently because it better captures the underlying kinematic manifold.
• Task-space (EEF) is competitive when data/compute is limited, and — critically — EEF becomes the superior choice under transfer learning (from π0) and cross-embodiment, where a body-agnostic SE(3) pose transfers across different kinematics while joint vectors do not.

5. The EEF-vs-Joint verdict (what you actually came for)

	Joint space (ℝⁿ)	Task / EEF space (SE(3))
Strength	Control stability; captures the kinematic manifold	Generalization / transfer
Scaling	Improves with more data/compute (esp. regression)	Strong in low-data / limited-compute
Transfer & cross-embodiment	Weaker (joint vectors are body-specific)	Superior (SE(3) pose is body-agnostic)
Risks	needs IK; singularities	FK is well-defined
Best when	one hardware platform, maximize performance	many bodies, transfer, generalization

And on the temporal axis, delta > absolute in essentially all standard settings — provided it is chunk-wise.

6. Practical guidelines (the paper's conclusion)

1. The chunking execution horizon k is not a constant — adapt it to the temporal abstraction (longer for absolute, shorter for delta).
2. Single-platform, resource-rich, maximize performance → joint space + chunk-wise delta is the most robust combination.
3. Generalized setting (cross-embodiment / transfer learning) → task space (EEF) is the superior spatial abstraction.

7. Significance

Most VLA papers inherit an action space without justifying it; this study turns that choice into a measured, reproducible decision and supplies a mechanistic reason (the O(k) vs O(1) noise-propagation result) for the single most common silent bug — step-wise delta chunking. It pairs naturally with VLA Architectures (the action-decoder axis) and the π-series choice of chunk-wise joint deltas, and it explains why cross-embodiment frameworks (Cross-Embodiment) gravitate to EEF/SE(3) action spaces. As an ICML-style "demystifying" study it is evidence, not a new model — but it is the reference to cite when defending an action-space choice.

No public code repository was found at review time; analysis is grounded in the arXiv paper (v1).

Links

• arXiv: 2602.23408
• ICML 2026: https://icml.cc/virtual/2026/poster/62805
• One-page summary: ICML-2026-Demystifying-Action-Space
• Figures 1–5 reproduced from the paper (Tsinghua IAIR, 2026) for scholarly review; © the authors.

Related pages

• VLA Architectures · Cross-Embodiment · VLA Attention
• ICML 2026 index

← Back to ICML-2026 · Home