Organizations: School of Artificial Intelligence, Beihang University, Beijing, China · School of Software, Beihang University, Beijing, China · School of Electronic and Information Engineering, Beihang University, Beijing, China · National College for Excellent Engineers, Beihang University, Beijing, China · Communication University of China, Beijing, China · Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, USA
Pretrained video representations provide rich temporal information for robotic manipulation, yet effectively integrating them into existing vision-language-action (VLA) policies remains challenging. We present Flow-Aware Visual Residual Alignment (FAVR), a lightweight framework for flow-aware visual conditioning of frozen VLA policies. FAVR maintains a compact temporal visual state through linear Test-Time Training (TTT) and retrieves action-conditioned visual features throughout flow-based action generation. By exploiting the affine structure of the frozen action head, we explicitly characterize how injected visual features influence predicted action velocities and formulate flow-aware channel selection as a box-constrained optimization problem. We develop a FISTA-based solver for joint gate optimization and derive an efficient closed-form approximation through diagonalization and rescaling. With only 1.74-2.19M trainable parameters and 5,000 adaptation steps, FAVR improves pretrained VLA policies across benchmarks. On RoboTwin 2.0, FAVR improves LingBot-VLA from 87.12% to 90.04% under clean conditions and from 86.88% to 89.38% under randomized conditions. A controlled ablation on five RoboTwin tasks further shows that flow-aware gating increases average success from 72.0% with a gate shared across flow steps to 82.0%, demonstrating the benefit of adapting visual conditioning throughout action generation.
Figures & tables
Figure 1: Shared versus flow-aware gating under the same video context. The heatmaps show recorded channel weights at representative flow steps. On five RoboTwin 2.0 tasks, average success increases from 72.0% with a shared gate to 82.0%.
Figure 2: Overview of FAVR. Observed video features update a linear TTT state queried by action features at each flow step. A predicted velocity residual guides FISTA-based joint or approximate closed-form gating before the frozen affine action head. Executed actions provide subsequent observations for state updates.
Figure 3: Comparison of joint and closed-form gate optimization. FISTA iteratively optimizes channel weights while retaining cross-channel coupling, whereas the closed-form approximation first computes a diagonal gate gD and then applies analytic rescaling to obtain gCF=α⋆gD .
Method
Camera
Robot
Language
Light
Background
Noise
Layout
Avg.
VLA-JEPA [ 32 ]
63.30
67.10
85.40
95.60
93.60
66.30
85.10
78.00
π0.5 (Frozen)
68.36
78.58
86.34
95.79
95.91
86.45
85.25
84.22
JEPA-VLA [ 25 ]
67.23
76.52
85.82
93.43
91.54
81.07
84.66
81.96
PVI [ 37 ]
69.17
77.48
87.83
95.18
92.19
81.07
85.31
83.10
FAVR-CF
69.17
79.10
86.92
96.76
95.35
85.63
85.90
84.55
FAVR
70.42
79.87
88.03
96.41
96.00
85.82
85.64
85.05
Table 1: Success rates (%) on LIBERO-Plus across seven perturbation categories. Avg. is computed over all 10,030 evaluation scenarios.
Method
Spatial
Object
Goal
Long
Avg.
OpenVLA-OFT [ 17 ]
97.6
98.4
97.9
94.5
97.10
MemoryVLA [ 31 ]
98.4
98.4
96.4
93.4
96.65
VLA-JEPA [ 32 ]
96.2
99.6
97.2
95.8
97.20
π0.5 (Frozen)
98.0
98.2
97.0
96.4
97.40
JEPA-VLA [ 25 ]
98.8
99.4
97.8
95.2
97.80
PVI [ 37 ]
98.4
100.0
97.2
94.6
97.55
Table 2: Success rates (%) on LIBERO.
Method
C2C
C2R
π0.5 [ 35 ]
82.74
76.76
LingBot-VLA
87.12
86.88
FAVR
90.04
89.38
Table 3: Success rates (%) on RoboTwin 2.0. C2C and C2R denote Clean-to-Clean and Clean-to-Randomized evaluation, respectively.
Figure 4: Representative Cobot Magic rollouts: block pick-and-place, paper-cup grasping, covered-block retrieval, and history-conditioned cup selection. The last two tasks require visual history and separate observation from execution.
Task
Initial FT
Continued FT
FAVR(Ours)
Block PnP
10/20
12/20
15/20
Paper-cup grasping
8/20
11/20
13/20
Table 4: Real-world manipulation results on Cobot Magic. Each task uses 35 demonstrations and 20 evaluation trials.
Method
Hanging mug
Can/pot
Bell
Bread/skillet
Can/basket
Avg.
Frozen baseline
47%
72%
77%
84%
64%
68.8%
Non-flow-aware gate
50%
66%
88%
88%
68%
72.0%
Learned gate
42%
74%
70%
90%
72%
69.6%
Direct output
50%
86%
90%
90%
78%
78.8%
Hidden-only
52%
82%
86%
92%
86%
79.6%
Window-wise reset
52%
86%
88%
90%
72%
77.6%
Table 5: Ablation results on five RoboTwin 2.0 tasks. All adapted variants are trained for 1K steps and evaluated over 50 trials per task. The frozen baseline uses previously obtained evaluation results.
Figure 5: Benefits of multi-view visual conditioning in open_microwave . Synchronized rollout frames show that gripper slippage is difficult to identify from the head camera but visible from the wrist camera.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Base policy
Steps
LR
Eff. batch
Adapter data
V-JEPA views
FISTA / horizon
LIBERO
π0.5
5,000
5×10−5
32
LIBERO demonstrations
1
8 / 50:10
LIBERO-Plus
π0.5
5,000
5×10−5
32
14,347 perturbed demos
1
128 / 50:10
RoboTwin 2.0
LingBot-VLA 4B
5,000
5×10−5
8
50 clean demos/task
3
128 / 50:50
Appendix
Table S1: Main benchmark training and evaluation configurations. Steps are optimizer updates and the evaluation horizon is written as predicted/executed actions. Views refer to the auxiliary V-JEPA branch; the original VLA observation interface is retained.
Policy
Spat.
Obj.
Goal
Long
Total
Frozen π0.5
490
491
485
482
1,948
FAVR-CF
496
497
488
484
1,965
FAVR
494
500
492
486
1,972
Appendix
Table S2: LIBERO success counts. Each suite has 500 trials (2,000 in total).
Perturbation
Base
FAVR
Δ (pp)
Camera viewpoint
68.36
70.42
+2.06
Robot initialization
78.58
79.87
+1.29
Language instruction
86.34
88.03
+1.69
Lighting
95.79
96.41
+0.62
Background texture
95.91
96.00
+0.09
Sensor noise
86.45
85.82
-0.63
Appendix
Table S3: LIBERO-Plus perturbation breakdown. Δ denotes FAVR minus the corresponding frozen π0.5 baseline, in percentage points; Avg. pools all 10,030 evaluation scenarios.
Task
π0.5 C2C
Frozen C2C
FAVR C2C
π0.5 C2R
Frozen C2R
FAVR C2R
adjust_bottle
100
100
100
99
98
99
beat_block_hammer
96
90
92
93
81
83
blocks_ranking_rgb
92
92
90
85
90
93
blocks_ranking_size
49
67
72
26
74
72
click_alarmclock
98
78
89
89
86
89
click_bell
99
77
92
66
91
94
Appendix
Table S4 : RoboTwin 2.0 success rates (%) for all 50 tasks. The π0.5 columns are reported in LingBot-VLA [ 35 ] (Table S9); Frozen LingBot-VLA and FAVR are our evaluations (100 trials per task per setting). C2C/C2R denote clean/randomized evaluation.
Position
Gate
Mug
Can/pot
Bell
Bread/skillet
Can/basket
Avg.
Layer 12
Head-derived
42
68
68
94
64
67.2
Layer 12
Learned
56
74
90
92
64
75.2
Layer 24
Head-derived
46
70
64
86
68
66.8
Layer 24
Learned
54
80
96
88
80
79.6
Layer 36
Head-derived
54
68
62
88
56
65.6
Layer 36
Learned
52
80
92
86
62
74.4
Appendix
Table S5: Five-task injection-position experiment. Success rates (%) for three earlier injection layers and the final affine interface. All variants are reported with per-task results.
Representation
Spatial
Object
Goal
Long
Avg.
I-JEPA
99.2
99.8
97.4
94.8
97.8
V-JEPA
98.8
100.0
98.4
97.2
98.6
Appendix
Table S6: Representation ablation on LIBERO: success rates (%). Both alternatives are used as frozen visual representations for the FAVR interface.
Task
Initial
Extra
Initial FT
Cont. FT
FAVR
Block pick-and-place
5K
3K
10/20
12/20
15/20
Paper-cup grasping
15K
5K
8/20
11/20
13/20
Appendix
Table S7: Standard real-world manipulation. Each task has 35 demonstrations. Initial fine-tuning establishes the shared task-adapted checkpoint; continued fine-tuning and FAVR use the same additional update budget. Outcomes are successes out of 20 trials.
Task
Demos
Steps
Correct / trials
Covered-block retrieval
35
5K
7/10
History-conditioned cup selection
35
5K
7/10
Appendix
Table S8: History-conditioned target selection on the real robot. Selection correctness is reported separately from full task completion.
Policy
Average latency (ms)
Frozen VLA baseline
1117.7
FAVR-CF
1134.3
FAVR
1487.3
Appendix
Table S9: Reported inference latency. Measurements are averages over 100 inference calls under the pipelined schedule described below, using RoboTwin 2.0 inputs on an NVIDIA RTX PRO 6000 Blackwell GPU. Each call predicts a 50-action chunk with ten flow steps.
Test-time adaptation of vision-language-action models helps robot policies adjust to new views, tasks, objects, and execution errors during deployment.
Autoregressive normalizing flows (NFs) provide a tractable density for learning and sampling continuous robot actions. We show that a single maximum-likelihood objective can train vision--language--action (VLA) policies competitive with leading methods on manipulation benchmarks. We present NF-VLA, which combines conditional autoregressive normalizing flows with score-based action refinement. Its deep--shallow architecture concentrates visual, language, and proprioceptive prefix conditioning in a high-capacity deep block, while unconditioned shallow blocks provide a shared invertible action transformation. The vision--language backbone and flow are jointly trained by exact maximum likelihood on Gaussian-smoothed action chunks. The same learned density supplies a single score-based correction at inference, refining generated actions to account for training-time smoothing without a separately trained denoiser. Trained on LIBERO from a pretrained vision--language backbone without robot pretraining, NF-VLA achieves 98.35% success across the four standard suites and 72.6% on LIBERO-Plus using the same checkpoint without further fine-tuning. Ablations show that deep-only prefix conditioning and score-based refinement both improve closed-loop success. Together, these results establish carefully structured autoregressive flows as a competitive approach to VLA control, with one learned density supporting policy training, action generation, and refinement.
Ruichen Deng, Qinyu Zhao, Yong-Hyun Park +3
Department of Computer and Information Science University of Pennsylvania · School of Computing Australian National University
Flow-based vision-language-action (VLA) policies offer strong expressivity for action generation, but suffer from a fundamental inefficiency: multi-step inference is required to recover action structure from uninformative Gaussian noise, leading to a poor efficiency-quality trade-off under real-time constraints. We address this issue by rethinking the role of the starting point in generative action modeling. Instead of shortening the sampling trajectory, we propose CF-VLA, a coarse-to-fine two-stage formulation that restructures action generation into a coarse initialization step that constructs an action-aware starting point, followed by a single-step local refinement that corrects residual errors. Concretely, the coarse stage learns a conditional posterior over endpoint velocity to transform Gaussian noise into a structured initialization, while the fine stage performs a fixed-time refinement from this initialization. To stabilize training, we introduce a stepwise strategy that first learns a controlled coarse predictor and then performs joint optimization. Experiments on CALVIN and LIBERO show that our method establishes a strong efficiency-performance frontier under low-NFE (Number of Function Evaluations) regimes: it consistently outperforms existing NFE=2 methods, matches or surpasses the NFE=10 π0.5 baseline on several metrics, reduces action sampling latency by 75.4%, and achieves the best average real-robot success rate of 83.0%, outperforming MIP by 19.5 points and π0.5 by 4.0 points. These results suggest that structured, coarse-to-fine generation enables both strong performance and efficient inference. Our code is available at https://github.com/EmbodiedAI-RoboTron/CF-VLA.
Fan Du, Feng Yan, Jianxiong Wu +8
1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology +1
Vision-language-action (VLA) policies based on flow matching generate action chunks through repeated evaluations of an action expert. Increasing the number of integration steps raises inference cost, but does not necessarily improve closed-loop success. We propose Coda, which reallocates part of this integration budget to a single learned endpoint correction. A frozen policy first completes a few-step noise-to-action trajectory; a lightweight Transformer then predicts a demonstration-supervised residual using the candidate action, source noise, and shared observation-prefix cache. Only the corrector is trained. On 50 RoboTwin Easy tasks, five-step Coda improves success from 71.64% to 74.68% over the matched five-step baseline, while reducing forward latency by 30.2% relative to the default ten-step policy. A two-step configuration achieves 71.88% success with a 2.12× speedup. An independent 13-task control shows a 5.69-percentage-point gain at nearly equal latency, supporting correction as an effective alternative to additional integration. The same design also improves frozen official SmolVLA, raising two-step success from 60.8% to 69.4%. These results show that endpoint correction improves the quality-latency trade-off of frozen flow-matching policies.
Zhipeng Tang, Xinda Chen, Weining Rao +5
Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · Zhongke Haichuan Intelligent