What happens when a robot policy's action numbers silently change meaning?

ActionShift freezes disjoint generalization splits over real ManiSkill tasks and re-maps what a policy's outputs mean — permutation, sign, scale, target, frame, lag, gripper — without telling the agent. The benchmark measures adaptation: 6 bounded probe steps, then success or failure. Task dynamics stay fixed, so every drop is attributable to the interface alone.

pip install actionshift   # 1.1.0 · Python 3.11+ · ManiSkill 3 tasks

One half of a pair

ActionShift is the online adaptation half of the hidden action-interface attack. Its sibling, ActionABI, is the offline forensic half: it recovers the same contract from logged trajectories and abstains when the evidence is thin. Both speak one contract grammar, and the coupling is real code — ActionABI's C++ evidence scorer runs live inside ActionShift's belief loop via pybind11, with verified numerical parity.

Evidence

PickCube adaptation ladder	No adaptation 0.000 → pool-belief (entropy probes) 0.987, privileged oracle 1.000 — same frozen PPO backbone, Wilson 95% intervals
Probe efficiency	DualABI reaches matched success with 52% fewer probe steps than the baselines
Contract identified in	6 bounded, safe probe steps — no retraining, never told the active contract
Full suite	320 tests incl. the simulator-free actionshift-selftest plug-and-verify run
Frozen splits	Hash-addressed seen/unseen contracts over PickCube, PushCube, PullCube, StackCube on ManiSkill 3
Open artifacts	Dataset + baseline checkpoints on HuggingFace · DOI 10.5281/zenodo.21500713

Media

Adaptation ladder on PickCube: 0.000 floor to 0.987 pool belief — Method ladder on PickCube/seen: floor 0.000, pool belief 0.987, oracle 1.000.

Probe budget vs success pareto — Probe budget vs success: DualABI's efficiency edge at matched success.

Latency identification under the lag side-by-side — Latency is part of the contract: identification under actuation delay.

Simulator-free selftest demos — The selftest demos: plug-and-verify without a simulator.

Install & verify

pip install -e '.[dev]'   # Python 3.11, see README for ManiSkill extras
actionshift-selftest     # simulator-free plug-and-verify run

Limits

What this release does not claim (from the claim boundary):

The safety mask is a software constraint check, not a hardware safety certificate.
Matched-privilege comparisons measure the value of information-seeking, not unprivileged adaptation.
No broad-novelty or method-superiority-over-SOTA claim — the graded ledger and red-team audit are in the README.

Family

Sibling	ActionABI — offline forensic recovery over the same contract grammar (repo: Archerkattri/actionabi)
Simulator	ManiSkill 3 (PickCube, PushCube, PullCube, StackCube)
More	Research portfolio · code · PyPI
External	ManiSkill docs