Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models

Shuaijun Liu1 Feiyang You1 Chengyu Wu1 Shuyang Hao1 Chenglong Zhang1 Jingyao Cai2 Xingwei Chen3,4 Ningxin Su1,*
1The Hong Kong University of Science and Technology (Guangzhou) 2National Centre for Computer Animation, Bournemouth University 3Shanghai Jiao Tong University 4Eastern Institute of Technology, Ningbo
*Corresponding author: ningxinsu@hkust-gz.edu.cn

Website: nebulis-lab.com/Beyond-Appearance-Shifts

Abstract

Vision-language-action (VLA) models have achieved strong performance in embodied manipulation, but still lack a clear mechanism to balance behavioral stability with task-semantic sensitivity. We identify two complementary failure modes. Under task-preserving changes, where task semantics remain unchanged but scene appearance varies (e.g., style, illumination, clutter, or paraphrasing), policies often exhibit unnecessary action drift. Conversely, under semantic-breaking changes, where key task semantics such as the target object or constraint are altered, policies frequently fail to produce sufficiently distinct behaviors and instead follow the original trajectory. To address this gap, we propose BAS-VLA, a task-semantic action calibration framework built on top of a frozen base VLA. BAS-VLA adopts a breaking-centered calibration core as the default path, and introduces a selective evidence-gated preserving auxiliary that activates only when nuisance variation is detected while task semantics remain consistent. On the OpenPI-pi0.5 / LIBERO-Object Milk-Swap benchmark, BAS-VLA maintains high success on clean (98.0%) and semantics-preserving conditions (97.5%), while reducing clean-criterion success to 0.0% under deliberate target-object swaps, demonstrating strong stale-task suppression and task-semantic separation. On validated style-preserving shifts, it improves success from 42% to 70% without degrading clean performance. These results highlight that reliable VLA behavior requires moving beyond appearance robustness toward explicit task-semantic action calibration.

Clean 98.0% success on Milk-Swap clean episodes
Control 97.5% success under preserving controls
Break 0.0% stale clean-task success under target swap
Style Shift +28.0 pts validated preserving-side gain
BAS-VLA overview from the paper

Overview

BAS-VLA frames reliability as an action-level calibration problem. Task-preserving interventions should remain in the same action basin, while semantic-breaking interventions should push the policy away from the stale clean trajectory.

  • Task-preserving changes: style, illumination, clutter, viewpoint, or wording changes that keep task semantics fixed.
  • Semantic-breaking changes: edits to target objects or constraints that require a different behavior.
  • Frozen carrier: the base VLA remains fixed; BAS-VLA attaches a lightweight calibration layer at the action output.

Figure 1. BAS-VLA overview. Part (A) summarizes the two intervention types studied in this work: task-preserving changes and semantic-breaking changes. Part (B) shows BAS-VLA as an action-calibration layer over a frozen base VLA. Part (C) illustrates the task-aware visual filtering probe, from RGB input to text-conditioned grounding, SAM 2 mask refinement, merged task-relevant masks, and the resulting probe image.

Motivation

Strong VLA backbones can still fail in opposite ways. Under appearance-only changes, they may drift from an otherwise correct behavior. Under true semantic changes, they may keep executing the original task. BAS-VLA separates these cases by making the calibration target explicit at the action level.

Task-preserving motivation triptych

Figure 2. Task-preserving motivation on a single Black-Bowl-to-Plate instruction. The three panels show a clean success, a failure under a darkened illumination shift, and a failure under a semantics-preserving noise perturbation.

Method

Breaking-Centered Core

The default BAS-VLA path adds a small residual calibrator to the frozen base action. Matched clean, preserving-control, and semantic-break triplets train the residual to keep clean/control actions close while pushing breaking edits outside the clean action basin.

Evidence-Gated Preserving Auxiliary

The preserving auxiliary is intentionally selective. It activates only when early rollout phase, frozen visual-feature discrepancy, semantic consistency, and action disagreement jointly indicate nuisance-induced drift rather than a true task change.

Evidence-gated preserving example

Figure 3. Dark-scene preserving example. The frozen carrier misreads the target under illumination change; the evidence-gated probe path activates and produces a corrected action.

Results

The tables below reproduce the corresponding paper tables in full rather than using shortened web summaries.

Table 1. Motivation evidence for reliability gaps before BAS-VLA. Original denotes native carrier performance on the unmodified task condition, and Perturbed denotes performance under the corresponding preserving or task-changed intervention.
CarrierSuitePerturbationOriginal success ↑Perturbed success ↑Δ (pts)
OpenVLA-OFTLIBERO-SpatialPerimeter clutter98/100 (98.0%)65/100 (65.0%)-33.0
OpenVLA-OFTLIBERO-SpatialMargin noise97/100 (97.0%)51/100 (51.0%)-46.0
OpenPI-pi0.5LIBERO-ObjectIllum. change98/100 (98.0%)89/100 (89.0%)-9.0
OpenPI-pi0.5LIBERO-ObjectTarget-object Swap197/200 (98.5%)68/200 (34.0%)-64.5
Table 3. Main task-preserving style-shift results on the Bowl-on-Ramekin style-shift setting. Each method uses 200 episodes over four seeds; Δstyle is measured relative to the baseline style-shift row.
MethodClean success ↑Matched-clean subset success ↑Style-shift success ↑Δstyle ↑
Baseline194/200 (97.0%)108/200 (54.0%)84/200 (42.0%)-
BAS-VLA195/200 (97.5%)108/200 (54.0%)140/200 (70.0%)+28.0 pts
Table 4. Main semantic-breaking results on matched target-object triplets. Each row uses 200 clean, 200 control, and 200 break episodes; lower is better in Break.
TaskSettingClean success ↑Control success ↑Break (clean) ↓Ctrl.-Break gap ↑
Canonical harder target-object tasks
BBQ-SwapBBQ sauce → ketchup180/200 (90.0%)185/200 (92.5%)0/200 (0.0%)92.5
Ketchup-Swapketchup → tomato sauce176/200 (88.0%)184/200 (92.0%)0/200 (0.0%)92.0
Milk-Swapmilk → orange juice196/200 (98.0%)195/200 (97.5%)0/200 (0.0%)97.5
Butter-Swapbutter → cream cheese188/200 (94.0%)182/200 (91.0%)113/200 (56.5%)34.5
OJ-Swaporange juice → milk197/200 (98.5%)193/200 (96.5%)0/200 (0.0%)96.5
Stronger interference in the orange-juice → milk setting
OJ-SwapHard noise184/200 (92.0%)185/200 (92.5%)0/200 (0.0%)92.5
OJ-SwapClutter+Noise192/200 (96.0%)183/200 (91.5%)0/200 (0.0%)91.5
Table 5. Matched-carrier external strategy comparison on Milk-Swap. Each strategy uses 200 clean, 200 control, and 200 break episodes; lower is better in Break.
StrategyClean success rate ↑Control success rate ↑Break (clean) ↓Ctrl.-Break gap ↑
BAS-VLA196/200 (98.0%)195/200 (97.5%)0/200 (0.0%)97.5
SGAC-AC (NeurIPS'25)194/200 (97.0%)195/200 (97.5%)2/200 (1.0%)96.5
AAC (CVPR'26)196/200 (98.0%)195/200 (97.5%)2/200 (1.0%)96.5
RTC (NeurIPS'25)197/200 (98.5%)192/200 (96.0%)4/200 (2.0%)94.0
PCD (ICLR'26)176/200 (88.0%)179/200 (89.5%)6/200 (3.0%)86.5
ST4-VLA (ICLR'26)189/200 (94.5%)190/200 (95.0%)45/200 (22.5%)72.5
HAMLET (ICLR'26)88/200 (44.0%)86/200 (43.0%)0/200 (0.0%)43.0
Table 6. Real-robot changed-task completion. Every case uses 80 trials per method, for 640 changed-condition rollouts in total. T1/T2 include the initial 17 trials per method.
Changed-task caseFrozen SR ↑BAS-VLA SR ↑Δ (pts)n / method
T1: block target19/80 (23.8%)47/80 (58.8%)+35.080
T2: fruit target8/80 (10.0%)31/80 (38.8%)+28.880
T3: receptacle target13/80 (16.3%)52/80 (65.0%)+48.880
T4: spatial relation11/80 (13.8%)36/80 (45.0%)+31.380
Aggregate51/320 (15.9%)166/320 (51.9%)+35.9320
Table 7. Changed-task outcome decomposition. New, Old, and Other/timeout are mutually exclusive; easy swaps aggregate Milk/OJ/BBQ/Ketchup.
Setting / methodNew ↑Old ↓Other/timeout ↓
Easy swaps / Frozen18/800 (2.3%)712/800 (89.0%)70/800 (8.8%)
Easy swaps / BAS-VLA548/800 (68.5%)0/800 (0.0%)252/800 (31.5%)
Easy swaps / norm-matched random residual12/800 (1.5%)708/800 (88.5%)80/800 (10.0%)
Butter → cream cheese / BAS-VLA53/200 (26.5%)113/200 (56.5%)34/200 (17.0%)
OJ → milk + visual interference / BAS-VLA127/200 (63.5%)0/200 (0.0%)73/200 (36.5%)
Table 8. External-protocol success without benchmark-specific tuning. LIBERO-Plus: four categories × ten cases × 120 rollouts per method. LIBERO-PRO: ten cases × 120 rollouts per method.
ProtocolBaseline ↑BAS-VLA ↑Δ (pts)
LIBERO-Plus2,632/4,800 (54.8%)3,741/4,800 (77.9%)+23.1
LIBERO-PRO443/1,200 (36.9%)857/1,200 (71.4%)+34.5
Initial real-robot clean/control and behavioral diagnostics
Table 11. Real-robot semantic-adaptation results on two deployment tasks. Each method-task pair uses 16 clean, 17 control, and 17 break episodes. Here Break reports success on the changed instruction, and the separation score averages old-target suppression and new-target first-commit.
TaskMethodn (c/ctrl/br)CleanControlBreakOld-target suppr.New-target first-commitSep. score
T1: group the specified colored block with its target set
baselineFrozen16/17/179/16 (56.3%)9/17 (52.9%)4/17 (23.5%)7/17 (41.2%)6/17 (35.3%)38.2
BAS-VLA defaultBAS-VLA16/17/1710/16 (62.5%)10/17 (58.8%)7/17 (41.2%)12/17 (70.6%)11/17 (64.7%)67.6
T2: place the specified fruit onto the plate
baselineFrozen16/17/178/16 (50.0%)8/17 (47.1%)3/17 (17.6%)6/17 (35.3%)5/17 (29.4%)32.4
BAS-VLA defaultBAS-VLA16/17/179/16 (56.3%)9/17 (52.9%)6/17 (35.3%)11/17 (64.7%)10/17 (58.8%)61.8

Image Gallery

The gallery below presents selected figures from the paper.

Citation

@misc{liu2026basvla,
  title     = {Beyond Appearance Shifts: Task-Semantic Action Calibration for VLA Models},
  author    = {Liu, Shuaijun and You, Feiyang and Wu, Chengyu and Hao, Shuyang and Zhang, Chenglong and Cai, Jingyao and Chen, Xingwei and Su, Ningxin},
  year      = {2026}
}