actionshift-baselines

ActionShift baselines

Reference checkpoints for ActionShift, a benchmark for manipulation policies adapting to a hidden action-interface contract. Code: github.com/Archerkattri/actionshift. Splits, eval configs, and the rollout logs these checkpoints produced: kattri15/actionshift.

These are reference implementations from one benchmark run, not claims of best-in-class performance. Exact numbers, intervals, and the honesty caveats for every method are in reports/*.md in the GitHub repo, in particular reports/gate1.md and reports/adaptation_tournament.md.

What's here

ppo/*_final_ckpt.pt            frozen Gate-1 competent PPO backbones, one per task
learned/*.pt                   unprivileged learned contract-identifier regressors
delay_aware/*.pt               delay-aware augmented-state PPO backbones (long_lag split)

ppo/ — the frozen base policies

The task backbone every adaptation method sits on top of. Official ManiSkill v3.0.1 PPO baseline, control mode pd_ee_delta_pose, trained per task, then frozen for the whole tournament. This is a fixed reference policy competent at the clean task, not a SOTA manipulation policy — it is deliberately unmodified so that the only variable across experiments is the action-interface contract, not the backbone.

Clean-task success (no hidden contract, i.e. identity): Pick and Push both 1.000 on seen. Under a hidden contract with no adaptation, all four collapse to the 0.000-0.007 floor — that collapse is the premise the rest of the benchmark measures adaptation against. PegInsertionSide-v1 has no frozen checkpoint here: it never cleared the Gate 0 competence floor at the official training budget and is excluded from the benchmark (see reports/peg_retry.md on GitHub).

learned/ — unprivileged learned contract identifiers

No pool, no grammar enumeration: these regressors learn to identify the active contract from calibrated response alone. Trained on 96 contracts / 34,560 env steps / 80 epochs (osi, osi_random_excitation, recurrent), or 96 contracts with a probe phase (probe_osi); held out on 12 contracts.

These are reported in the paper as a load-bearing negative result: on PickCube/seen, the passive learned identifier scores 0.000 end-to-end, the same floor as no adaptation at all — active probing against a declared hypothesis space (the belief-family methods, not included here as checkpoints, see below) is not something these unprivileged learners recover on their own. Do not read these as "the adaptation methods" — read them as the specific negative result the paper reports.

delay_aware/ — delay-aware augmented-state PPO (the long_lag split)

A new task backbone, not the frozen Gate-1 PPO above. Same official PPO hyperparameters, plus two changes: the observation is augmented with the last 4 canonical actions (K=4, restores the Markov property under delay), and training uses a per-episode action lag resampled from {0, 1, 2, 4}. This is the standard state-augmentation reduction for delayed MDPs (Katsikopoulos & Engelbrecht 2003), not a reproduction of DCAC or D-TRPO.

The randomized-lag checkpoint is the headline long-lag method (frozen oracle scores 0.027 Pick / 0.153 Push on the same split, so this is roughly 20x / 2.7x). The curriculum variant trades long-lag success for near-perfect instantaneous competence and is included because the paper documents both, not because it is the recommended checkpoint for long_lag evaluation.

What's not here

Belief-family methods (exact belief, entropy probing, fixed/random probes, DualABI and its ablations) have no checkpoint in this repo because they are analytic, not learned — they maintain a Bayesian posterior over the declared contract pool (configs/contracts/core.yaml in the dataset repo) and are fully specified by their config files (configs/method/dualabi*.yaml, configs/method/*probes*.yaml in the dataset repo), with no trained weights to publish. This is also true of the scale-corrector and grasp-channel adaptation variants. Diffusion Policy imitation backbones used for the brittleness study (Round 12 in the paper) exist locally but are not uploaded here; ask if you need them.

Intended use

Reproducing or extending the ActionShift tournament: load a ppo/ checkpoint as the frozen backbone, evaluate it through the hidden-contract wrapper on the splits in kattri15/actionshift, and compare an adaptation method's success against the oracle ceiling and no-adapt floor those splits report. The learned/ and delay_aware/ checkpoints are reference points on that ladder, not standalone manipulation policies to deploy.

Honest limits

• Sim-only, no hardware. All checkpoints were trained and evaluated in ManiSkill (SAPIEN/PhysX). No real-robot claim is made or implied.
• Reference baselines, not SOTA. The PPO backbones are deliberately plain official-baseline policies so the interface is the only variable; they are not tuned for maximum task performance. The learned/ regressors are reported specifically as a negative result in the paper.
• Method coverage is partial. Only oracle, no-adapt, the four learned identifiers, and delay-aware control have trained checkpoints wired to the frozen backbones as of this release. Domain-randomized, RMA-style, and the belief-family methods have config files but not trained-weight checkpoints (see above) — that gap is disclosed in reports/gate1.md on GitHub as a structured exclusion, not silently omitted.
• delay_aware/ checkpoints are separate backbones, not drop-in replacements for ppo/ on splits other than long_lag — do not benchmark them on seen and compare directly to the ppo/ row without reading the caveats in reports/adaptation_delay_aware.md.

License

MIT, same as the code and paper. See LICENSE in the GitHub repo.

Citation

@software{attri2026actionshift,
  author = {Attri, Krishi},
  title = {ActionShift: Hidden compositional action-interface adaptation benchmark},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.21500713},
  url = {https://github.com/Archerkattri/actionshift}
}

Zenodo DOI: 10.5281/zenodo.21500713. If you use the ManiSkill baselines these checkpoints derive from, cite ManiSkill separately.