Escape Is Not Success: Diagnosing Spatial-OOD Failures of Robot Action Interfaces

Abstract

Success conflates policies that stay put with those that approach but fail to lift. We score escape, distance-normalised approach and strict completion in simulated cube lifting (targets 5–25 cm radially and 10–30° angularly beyond demonstrations; three training seeds per cell). Three ACT policies complete 0, 0 and 1 of 70 radial episodes per seed yet differ: one stays within 1 cm of the demonstrated region, one escapes but rarely approaches, and one often approaches but rarely completes. Absolute arms approach in at most 4 of 70 radial episodes per seed in either frame and, for EEF commands, with every camera tested; residual arms in 27–60. Under residual EEF commands, a wrist camera adds 19–32 strict radial completions per seed; absolute policies rarely approach, so their zero completions cannot isolate camera effects. Per-seed stages and matched contrasts expose interface failure modes.

Setup

A Franka Panda lifts a 5 cm cube in Isaac Lab. A scripted expert gives 15 demonstrations at each of 28 base-frame positions: r ∈ {0.35, 0.40, 0.45, 0.50} m and θ ∈ {−30°, −20°, …, 30°}. The demonstrated envelope is the sector r ∈ [0.325, 0.525] m, |θ| ≤ 35°, half a lattice step past the outer demonstrations so grasps there are not escapes. Test objects sit 5–25 cm beyond the envelope radially and 10–30° beyond it angularly. Each position has two rollouts of up to 600 steps at 30 Hz, with shared position and episode seeds.

The evaluation scores three stages independently:

1. Escape. The grasp point's maximum excursion beyond the demonstrated envelope is strictly greater than 1 cm.
2. Approach. Distance-normalised progress toward the object's nominal spawn position is ≥0.25. Approach is 1 minus the closest planar grasp-point distance to the spawn position divided by the object's distance from the envelope, clipped to [0, 1]: 0 if the grasp point never got nearer the object than the envelope edge, and 1 if it passed directly over it.
3. Strict completion. The cube's centre stays at least 1 cm above its rest height and within 6 cm of the grasp point in the plane at three consecutive checks, 15 control steps apart.

Results

Similar success, different stages. Three ACT policies complete 0, 0 and 1 of 70 radial episodes per seed: joint absolute never leaves the envelope by more than 1 cm; wrist-only absolute EEF leaves it in 40–44 episodes but approaches in only 2–4; per-query residual joint leaves in 49–52 and approaches in 29–31.

The arm's command interface changes approach. In the main 2×2 comparison, no absolute-arm radial episode closes a tenth of the gap; residual cells average 0.31–0.69 approach and complete 3–10 of 70 episodes. All four cells use a head camera and absolute gripper commands.

Counts transcribe the camera-ready Table 1. Absolute EEF uses quaternion orientation; residual EEF uses a rotation vector. A separate absolute rotation-vector control completes 55–56 of 56 training-cell episodes per seed, but no test episode reaches 0.25 approach. Closed-loop re-evaluation is not bitwise reproducible, so per-seed stage and completion counts can differ slightly from the paper.

Under residual EEF commands, strict radial completions rise from 6/5/1 with head views to 25/37/31 with head+wrist views and 27/24/32 with wrist views; absolute EEF approaches in only 0–4 of 70 radial episodes per seed, so its zero completions cannot isolate camera effects.

Recommendation

The study tests one simulated task, object, robot, top-down expert and ACT recipe; completion checks height and proximity, not contact; approach varies by seed. For spatial-OOD reaching, report:

1. Demonstrated region, envelope construction, target distances, and all stage thresholds.
2. Command frame, absolute or residual action, anchor, gripper, rates, and cameras.
3. Per-seed continuous escape/approach and thresholded stage and completion counts.
4. Count stage-order violations separately with values and completion flags (the moved-cube completion: a = 0.23 < 0.25).

BibTeX

@inproceedings{kim2026escape,
  title = {Escape Is Not Success: Diagnosing Spatial-{OOD} Failures of Robot Action Interfaces},
  author = {Kim, Juno and Park, Yesol and Yoon, Hye-Jung and Zhang, Byoung-Tak},
  booktitle = {CoRL 2026 Workshop on Everything Beneath the Policy},
  year = {2026},
  url = {https://openreview.net/forum?id=nRU0GzU0Hc}
}