Vision-language-action (VLA) models provide a scalable path toward generalist robotic manipulation by integrating visual perception, language understanding, and continuous action control. However, we reveal a critical limitation of VLA architectures: the action expert has limited access to the 3D geometric and 2D semantic information available in VLM features. This accessibility gap weakens perceptual grounding and limits performance on fine-grained robotic manipulation. To address this issue, we propose V-Link, which explicitly recovers visual representations during the vision-language (VL) to action (A) feature transfer. Specifically, V-Link learns complementary Spatial and Semantic Query representations within the VLM and injects them into Action DiT through asymmetric pathways. Semantic Queries complement the original VLM image tokens, whereas Spatial Queries provide dedicated geometric conditioning for spatially grounded action generation. Across LIBERO, LIBERO-Plus, and RoboTwin 2.0, our V-Link improves the average success rate over base model GR00T N1.6 by +1.9%, +31.2%, and +18.8%, respectively. On the AGIBOT A3 Ultra, V-Link further achieves gains of +20% and +24% on two real-world humanoid tasks.
Figures & tables
Fig. 1 : Key insight. GR00T N1.6 exhibits a visual representation accessibility bottleneck during VL-to-A feature transfer. Our V-Link addresses this bottleneck by learning complementary Spatial and Semantic Query representations and injecting them into Action DiT. The resulting Action DiT features support substantially better depth estimation and semantic segmentation, providing qualitative evidence of visual representation recovery.
Fig. 2 : Diagnostic evaluation protocol. We freeze the VLM and Action DiT features and train only lightweight task heads for depth estimation and semantic segmentation. Test-set performance serves as a controlled measure of visual representation accessibility. Left: Compared with VLM features, GR00T N1.6 Action DiT features exhibit a pronounced accessibility gap. Right: V-Link substantially narrows this gap in Action DiT on both diagnostic tasks.
Fig. 3 : The overall pipeline of V-Link. V-Link introduces learnable Spatial and Semantic Queries into the pretrained VLM. A causal attention mask enables them to independently extract geometric and semantic cues from deep VLM features without disrupting the native multimodal token stream. Lightweight depth and semantic heads provide explicit supervision to disentangle these cues within the VLM latent space and are removed at inference. The resulting representations are injected into Action DiT through asymmetric pathways: Semantic Queries complement the VLM image tokens, whereas Spatial Queries form a mandatory geometry-conditioning branch that compels Action DiT to explicitly encode depth-aware representations rather than rely on semantic shortcuts.
Fig. 4 : t-SNE on RoboTwin2.0 rollouts. We randomly sample 5,000 observations and extract the final-layer VLM outputs for image tokens, Spatial Queries, and Semantic Queries. Each token set is mean-pooled before projection. The separated distributions qualitatively indicate that the two query sets learn distinct representations.
Method
Spatial
Object
Goal
Long
Average
2D VLA
OpenVLA [ 21 ]
84.7
88.4
79.2
53.7
76.5
Dita [ 22 ]
84.2
96.3
85.4
63.8
82.4
CoT-VLA [ 23 ]
87.5
91.6
87.6
69.0
83.9
π0 -FAST [ 9 ]
96.4
96.8
88.6
60.2
85.5
π0 [ 24 ]
96.8
98.8
95.8
85.2
94.2
TABLE I : Comparison on LIBERO benchmark.
Method
Camera
Robot
Language
Light
Background
Noise
Layout
Total
OpenVLA-OFT [ 25 ]
56.4
31.9
79.5
88.7
93.3
75.8
74.2
69.6
π0 [ 24 ]
13.8
6.0
58.8
85.0
81.4
79.0
68.9
53.6
π0 -Fast [ 9 ]
65.1
21.6
61.0
73.2
73.2
74.4
68.8
61.6
Evo-Depth [ 28 ]
47.2
49.2
78.9
88.1
76.4
77.6
69.6
69.6
GR00T-N1.6 (Base)
21.2
40.3
35.5
65.1
76.7
27.4
51.0
43.8
V-Link (Ours)
42.8
59.1
82.8
95.7
94.9
87.8
74.4
75.0
TABLE II : Comparison on LIBERO-Plus benchmark.
Method
Move
Place
Contact
Average
Pillbottle
Stapler
Bread Skillet
Phone Stand
Press Stapler
Turn Switch
π0 [ 24 ]
21
0
23
35
62
27
28.0
Xiaomi Robotics-0 [ 29 ]
65
16
63
44
79
38
50.8
EventVLA [ 30 ]
60
14
64
64
80
24
51.0
π0.5 [ 7 ]
46
20
76
65
79
52
56.3
GR00T N1.6 (Base)
58
7
23
54
80
6
38.0
TABLE III : Comparison on Robotwin2.0 benchmark.
Method
Move
Place
Contact
Average
GR00T N1.6
32.5
38.5
43.0
38.0
+ Spatial Query
44.5
54.0
55.0
51.2
+ Semantic Query
35.5
47.0
61.0
47.8
+ Both Query
48.0
57.5
65.0
56.8
TABLE IV : Ablation study of auxiliary queries.
Method
Move
Place
Contact
Average
GR00T N1.6
32.5
38.5
43.0
38.0
w/o Task Head
35.5
38.5
47.0
40.3
w/ Task Head
48.0
57.5
65.0
56.8
TABLE V : Ablation study of task supervision. “w/o Task Head” retains the Spatial and Semantic Queries and their injection mechanisms, while removing the depth and semantic heads and the ground-truth supervision.
Method
Move
Place
Contact
Average
GR00T N1.6
32.5
38.5
43.0
38.0
w/o Injection
40.5
28.5
50.5
39.8
w/ Injection
48.0
57.5
65.0
56.8
TABLE VI : Ablation study of auxiliary query injection. “w/o Injection” retains the Spatial and Semantic Queries, the depth and semantic heads, and the ground-truth supervision, but does not inject both queries into Action DiT.
Fig. 5 : Qualitative comparison. The Action DiT of GR00T N1.6 exhibits severe visual representation degradation, retaining little 3D depth information and incomplete 2D semantics. V-Link effectively recovers the spatial and semantic representations within Action DiT.
Fig. 6 : Diagnostic evaluation of frozen visual representations. The poor results with random features show that diagnostic performance depends on the information encoded in the frozen features, rather than the task heads themselves. Spatial and Semantic Queries achieve the best depth-estimation and semantic-segmentation results, respectively, demonstrating task-specific specialization. V-Link Action DiT outperforms GR00T Action DiT on both tasks, confirming successful representation recovery.
Fig. 7 : Effect of Spatial/Semantic Query token number on performance and inference latency. Latency is tested on a single H100 and averaged over 1,000 randomly sampled inputs with batch size 1, using torch.compile, 4 denoising steps, and 3 views input.
Fig. 8 : Real-robot evaluation on the AGIBOT A3 Ultra. Over 50 consecutive trials per task, V-Link achieves 98%/94% success rate on autonomous power-on/off tasks, versus 78%/70% for base model GR00T N1.6.
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · The Chinese University of Hong Kong · Simplexity Robotics