Anonymous ICRA 2027 Submission

ASRR

Offline Action Sequence Residual Refinement for Robot Policy Adaptation

Treat the action proposal as an editable adaptation interface. ASRR keeps the robot policy frozen and learns a compact residual refiner from cached outputs and existing demonstrations.

Paper Code Videos Interactive Demo
π0.5 · LIBERO-10Base
π0.5 · LIBERO-10ASRR
Corn to plate · 3×Base
Corn to plate · 3×ASRR

Interactive explainer

Follow one proposal through ASRR

Follow the local error, editable interface, residual correction, and recorded evidence in 1 minute 45 seconds.

Open full screen
ASRRAction sequence refinement, step by step

Illustrative action-space projection

A useful proposal with one local error

Illustrative
Base and refined action trajectories A base action sequence approaches a target with a local error. A bounded residual moves the selected action toward a refined trajectory.
00:00 / 01:45

Recorded execution

Evidence comparison

Select an evidence case to load its recorded Base and ASRR rollouts.

ASRR explainer ready.

Abstract

Improve the proposal, keep the policy fixed

Robots deployed with limited local computation and intermittent remote support need efficient ways to improve an existing policy. Such a policy may already capture task intent and propose reasonable actions, yet local inaccuracies can still cause failures. Further fine-tuning or reinforcement learning can require costly optimization or additional interaction.

Action Sequence Residual Refinement (ASRR) tests whether a useful action proposal can itself serve as the target of adaptation. It trains a lightweight refiner offline from cached policy outputs and demonstrations, optionally using policy-side context. Across action-sequence policies, modern VLAs, and three real-robot tasks, the results establish predicted action sequences as editable adaptation interfaces.

Method

Frozen proposal, optional context, residual edit

ASRR predicts a conservative residual in the policy-native action space. The corrected proposal follows the original execution rule, while training reuses cached outputs and existing demonstrations.

ASRR method overview, offline refiner training, and qualitative refinement examples
ASRR-A uses the proposal alone; ASRR-C adds context exposed by the frozen policy.

Frozen base

The selected policy checkpoint is unchanged during refiner training.

Sequence editing

The full temporally structured proposal is the object being corrected.

Offline fitting

Cached proposals decouple refiner optimization from large-policy inference.

Native execution

The refined sequence retains the controller's action horizon and replanning rule.

Results

From action chunks to VLAs and real robots

All reported refiners keep their corresponding base policy frozen.

π0.5 · four-suite mean+14.4 ppContext-conditioned, suite-specific refinement
5k checkpoint62.0 → 80.9Mean LIBERO success (%)
Recorded compute25–63% lessGPU-hours than adjacent continued LoRA runs
Real robot · aggregate+6.7 pp154/240 to 170/240 successes
Table I from the paper: simulation results for ACT, Diffusion Policy, pi0.5 suite-specific and shared refiners, and OpenVLA-OFT
Table I from the paper. Δ is percentage-point gain; parameters are base/refiner millions.

Analysis

When refinement helps, and what it costs

The same-run checkpoint study isolates base competence, while the context comparison examines what information the refiner needs.

ACT success across refinement horizons under direct chunk execution and temporal aggregation
ACT execution rules favor different editable horizons.
Pi0.5 base competence and action-only versus context-conditioned refinement comparisons
Earlier capable checkpoints gain more; context improves the tested π0.5 refiners.

Recorded adaptation compute

From the same starting checkpoint, cached ASRR fitting uses fewer recorded GPU-hours than continuing LoRA to the adjacent checkpoint. Cache extraction is the dominant one-time cost and can be reused for compatible refiner variants at the same checkpoint and dataset.

Checkpoint pair5k→10k10k→20k20k→30k
Continued LoRA (GPU-h)15.7931.5131.18
ASRR cache + fitting (GPU-h)11.8111.5511.52
Compute reduction25.2%63.3%63.1%
Success: start→ASRR / later Base62.0→80.9 / 74.674.6→89.1 / 91.291.2→96.4 / 91.3

Simulation videos

Selected action-sequence rescues

Representative Base failures and ASRR successes from policy families evaluated in the paper.

π0.5 · LIBERO-10

Base failure
ASRR success

OpenVLA-OFT · Goal

Base failure
ASRR success

Real-robot evaluation

Three Piper manipulation tasks

Each method is evaluated over 80 trials per task with the same frozen base policy. Videos play at 3× recorded speed.

Corn-to-plate

58/80 → 65/80
3×
Base failure
3×
ASRR success

Block-to-holder

53/80 → 57/80
3×
Base failure
3×
ASRR success

Block stacking

43/80 → 48/80
3×
Base failure
3×
ASRR success

Implementation

Policy-agnostic refiner core

The repository provides the shared ASRR modules and integration examples for action-sequence policies and VLA controllers evaluated in the paper.

CoreExamples
from asrr_core import ActionSequenceResidualAdapter, apply_residual

refiner = ActionSequenceResidualAdapter(
    action_dim=7,
    horizon=16,
    fusion_mode="action_only",
    head_type="bounded_dense",
)

delta = refiner(base_action)
refined = apply_residual(base_action, delta, alpha=1.0)

Citation

Anonymous submission

@misc{asrr2026,
  title = {ASRR: Offline Action Sequence Residual Refinement
           for Robot Policy Adaptation},
  author = {Anonymous},
  year = {2026},
  note = {Anonymous ICRA 2027 submission}
}