Understanding three-dimensional physical structure has become essential for vision-language-action models (VLAs). However, existing approaches either rely on explicit geometric inputs at inference time or adopt a single visual pathway, potentially compromising the visual-semantic knowledge inherited from pre-trained models. We present SPARK-VLA, Spatial Representation distillation and Knowledge routing for VLAs, a preserve-specialize-route framework for spatially adapting VLAs without requiring external 3D inputs or experts at inference. SPARK-VLA consists of two key components: (1) spatial representation distillation, which transfers 3D understanding from a frozen 3D expert to a LoRA-adapted visual branch by aligning attention responses while preserving the original visual pathway, and (2) a lightweight knowledge router, which selectively forwards action-relevant visual tokens to the language model using routing supervision derived from action-loss gradients. Across LIBERO, LIBERO-Plus, CALVIN, and real-world manipulation, SPARK-VLA achieves strong performance with only a 0.5B language backbone and approximately 53M trainable parameters, providing an effective and parameter-efficient approach to 3D-aware VLA learning. Code is available at https://github.com/park-jungin/spark-vla.
Figures & tables
Figure 1: We depict (a) the averaged success rate for LIBERO and (b) the average length for CALVIN, with respect to the total number of parameters. The center point and the size of circles indicates the averaged performance and the size of LLMs in each model, respectively.
Figure 2: Overview of SPARK-VLA. Trainable modules are represented in the red boxes.
Method
Spatial
Object
Goal
Long
Average
LLM Size
Train. Params.
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
Large-scale models ≥ 7B
Octo ( Ghosh et al., 2024 )
78.9
85.7
84.6
51.1
75.1
-
93M
OpenVLA ( Kim et al., 2024 )
84.9
88.4
79.2
53.7
76.5
7B
111M
OpenVLA-OFT ( Kim et al., 2025 )
97.6
98.4
97.9
94.5
97.1
7B
111M
Spatial Forcing ( Li et al., 2025a )
99.4
99.6
98.8
96.0
98.5
7B
291M
Table 1: Comparisons with state-of-the-art methods on LIBERO. * incorporate extra depth or point cloud information from other sensors. Bold indicates the best performances among tiny-scale models.
Method
Suite
Camera
Robot
Language
Light
Background
Noise
Layout
Total
LLM Size
Train. Params.
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
Large-scale models ≥ 7B
OpenVLA ( Kim et al., 2024 )
Avg.
0.8
3.5
23.0
8.1
34.8
15.2
28.5
15.6
7B
111M
OpenVLA-OFT ( Kim et al., 2025 )
Avg.
56.4
31.9
79.5
88.7
93.3
75.8
74.2
69.6
7B
111M
UniVLA ( Wang et al., 2026b )
Avg.
1.8
46.2
69.6
69.0
81.0
21.2
31.9
42.9
7B
8B
1B ≤ Small-scale models < 7B
Table 2: Comparisons with state-of-the-art methods on LIBERO-Plus. Bold indicates the best performances.
Method
Task
Tasks Completed in a Row
Avg. Len ↑
LLM Size
Train. Params.
1
2
3
4
5
MCIL ( Lynch and Sermanet, 2020 )
ABCD → D
0.377
0.027
0.002
0.000
0.000
0.40
-
-
RT-1 ( Brohan et al., 2022 )
ABCD → D
0.844
0.617
0.438
0.323
0.227
2.45
-
-
Robo-Flamingo ( Li et al., 2024 )
ABCD → D
0.964
0.896
0.824
0.740
0.660
4.09
7B
1.4B
GR-1 ( Wu et al., 2024 )
ABCD → D
0.949
0.896
0.844
0.789
0.731
4.21
-
-
UP-VLA ( Zhang et al., 2025b )
ABCD → D
0.962
0.921
0.879
0.842
0.812
4.42
-
1.5B
Table 3: Comparisons with state-of-the-art methods on CALVIN benchmark. Bold indicates the best performance.
Figure 3: Overview of our real-world robot setup and tasks. We use a 7-DoF robotic arm with two cameras for manipulation. We evaluate six distinct tasks across a clean environment and a more complex kitchen environment.
Figure 4: Performance comparison on real-world tasks.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Value
GPUs
NVIDIA RTX A6000 48GB × 4
Batch size per GPU
16
Initial learning rate / Final learning rate
5e-5 / 5e-8
Warm-up with Equation 13
2k steps
Learning rate scheduler
Cosine annealing lr scheduler
Input images
1 exo-view + 1 wrist view
Appendix
Table A.1: Hyperparameters for simulated experiments.
Figure A.1: Performance according to the action generation objective. NTP and Diffusion denote next token prediction and diffusion policy, respectively.
Method
Spatial
Object
Goal
Long
Average
Train. Params.
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
(a) only LLM LoRAs
-
-
-
-
< 5%
31.4M
(b) + Vision LoRAs
-
-
-
-
< 5%
60.1M
(c) + Projector
82.6
84.4
77.4
45.0
72.4
64.9M
(d) - SigLIP
62.4
76.8
58.4
37.2
58.7
48.8M
(e) + Attention Distill.
95.8
96.0
94.0
90.6
94.1
48.8M
Appendix
Table A.2: Ablation study for each component in our SPARK-VLA.
Spatial supervision
Success rate (%)
Action loss only
81.6
Feature matching
94.8
Relational matching
95.6
Attention-map matching
96.4
SPARK-VLA (ours)
98.0
Appendix
Table A.3 : Comparison with different spatial supervision strategies on LIBERO-Goal.
Fusion strategy
Success rate (%)
Direct all-token concatenation
94.0
FiLM-based branch fusion
94.8
Action-driven knowledge router (ours)
98.0
Appendix
Table A.4 : Comparison with different knowledge fusion strategies on LIBERO-Goal.
Momentum
0.0
0.9
0.99
0.999
0.9999
Avg. SR (%)
94.0
98.0
97.6
97.2
97.0
Appendix
Table A.5: Performance according to the momentum parameter β on LIBERO-goal.
Figure A.2 : Ablation study for each hyperparameter. We report the performance evaluated on LIBERO-Object according to the values of (a) the budget coefficient λbud , (b) the entropy coefficient λent , and the target keeping ratio ρ .
Method
Suite
Camera
Robot
Language
Light
Background
Noise
Layout
Total
LLM Size
Train. Params.
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
SR (%)
Large-scale models ≥ 7B
OpenVLA-OFT ( Kim et al., 2025 )
Spatial
88.3
40.0
80.5
98.3
97.3
96.3
93.9
84.0
7B
111M
Object
38.9
25.4
99.0
73.7
97.6
72.3
71.8
66.5
Goal
62.0
25.2
53.2
93.9
92.5
75.2
59.1
63.0
Long
38.7
38.2
87.0
89.4
86.8
63.5
76.9
66.4
Appendix
Table A.6 : Per-task and per-perturbation performance evaluated on LIBERO-Plus. Bold indicates the best performances.
Figure A.3 : Qualitative analysis of learned spatial representations. From given exocentric and wrist view observations, we extract depth maps using the original VGGT and the spatial path in our SPARK-VLA.
Figure A.4 : Qualitative analysis of selected tokens. From given exocentric and wrist view observations, the knowledge router in our SPARK-VLA selectively forwards instruction- and action-relevant visual (2D) and spatial (3D) tokens to the LLM. We mask out the tokens, which are not selected in the router.
Figure B.5 : Qualitative experiments of our method in real-world tasks.
Figure B.6 : Success rates on four additional kitchen manipulation tasks (left) and additional real-world examples (right). The success-rate plot reports 10 trials per task per method; Avg. denotes the mean across the four tasks.
Object
Condition
SPARK-VLA
w/o VGGT
Red cup
Training object
80%
60%
Blue cup
Unseen color
40%
40%
Black cup
Unseen color
60%
40%
Green handleless paper cup
Unseen color and geometry
40%
40%
Small milk carton
Unseen category and geometry
60%
60%
Hand-cream bottle
Unseen category and geometry
80%
40%
Appendix
Table B.7 : Pick-and-place success rates for SPARK-VLA and SPARK w/o VGGT on the training red cup and five held-out objects. Each object is tested at the same five predefined starting locations (five trials per object).
Language instruction
SPARK-VLA
w/o VGGT
“Pick the cup in the sink.”
Complete
Complete
“Move the cup to the sink.”
Complete
Failure
“Pick the cup and place it in the right-side area.”
Complete
Failure
“Move the red object to the right-side area.”
Partial
Complete
“Get the cup ready to wash in the sink.”
Partial
Failure
“Pick up the handleless green cup and put it in the sink.”
Complete
Complete
Appendix
Table B.8 : Outcomes for SPARK-VLA and SPARK w/o VGGT on nine language instructions held out from task training. Complete success requires executing the requested manipulation with the correct target; partial success denotes a correct grasp without completing the requested placement.
Recent Vision-Language-Action (VLA) methods improve generalization by aligning their representations with 3D scene geometry. However, these methods are fundamentally instruction-agnostic: the representations align the entire scene uniformly, neglecting the 3D geometry of the specific target object designated by the language instruction. This causes failures on fine-grained manipulation and target occlusion tasks, where success depends on accurate 3D understanding of the target object rather than the entire scene. To address this, we present Mind-VLA, an instruction-aware spatial representation alignment method for VLA models. Specifically, Mind-VLA first obtains the target object specified by the language instruction, then prepares its target-object tri-view and extracts the corresponding VAE and VGGT features. Finally, the latent representation of the VLA model is aligned with these features to enable instruction-aware 3D understanding. Mind-VLA reaches 93.9% on LIBERO and 4.47 on CALVIN with a compact 345M-parameter backbone. On real-robot tasks with target occlusion, Mind-VLA reaches 54% average success, outperforming the best-performing instruction-agnostic method in real-robot comparison by 32 percentage points. Code will be publicly available.
Xingyu Ding, Yuzhong Zhao, Yang Wu +4
Nanjing University, Nanjing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China
Vision-Language-Action models have achieved remarkable progress in robotic manipulation, yet they suffer from a critical limitation: a lack of 3D scene understanding. This deficiency manifests as three intertwined challenges: weak extraction of 3D spatial positions without enforcing multi-view consistency, inadequate 3D instance understanding, and fragile reasoning under occlusion. Although mature 3D perception methods exist, their direct integration into VLA pipelines is hindered by architectural incompatibility and by heavy reliance on costly instance-level annotations. To address the above challenges, we propose 3DVLA, a plug-and-play framework that injects robust 3D reasoning into pretrained VLAs without requiring extra manual labels or discarding VLM priors. Specifically, 3DVLA tackles the three challenges through: (1) pervasive 3D feature encoding with explicit multi-view consistency constraints across all modalities and a Spatially-Conditioned Geometry Aggregation method, (2) an instance estimation module with high-level instance tokens for 3D instance awareness, and (3) a masked self-supervised 3D encoding branch that retains its predictor for visual token completion to handle occlusions. We integrate 3DVLA with multiple VLA baselines and evaluate on LIBERO-Plus and RoboTwin 2.0. Results show consistent and significant gains in manipulation performance, validating both the effectiveness and plug-and-play compatibility of our approach.
Zhongyu Xia, Yousen Tang, Bingqing Wei +1
Wangxuan Institute of Computer Technology, Peking University · This work was done as an intern at PKU.
We propose a 3D-thinking-guided co-training framework that enables vision-language-action (VLA) models to perform 3D spatial reasoning implicitly during action prediction. Our core insight is that 3D geometry perception and 3D spatial reasoning are distinct capabilities that can be disentangled and injected at different feature hierarchies. During training, three tightly coupled components work in concert primarily within the latent space: (1) To gain geometric priors, a latent 3D geometry perception module aligns intermediate visual features with a 3D foundation model, acquiring low-level geometric cues without architectural modifications to the VLM backbone. (2) Complementing this, an online 3D reasoning distillation module mitigates the prompt-induced reasoning gap via a shared reasoning anchor token. During 3D VLM co-training, this anchor is emitted as the first output token to robustly encode spatial priors. During VLA training, it serves as an input token inserted between the task and action instructions, transferring high-level spatial thinking from explicit teacher reasoning prompts to student action prompts without chain-of-thought text generation. (3) These disentangled geometric and reasoning features are then united by a spatially augmented action integration, which jointly injects them into the action-query tokens as hierarchical spatial conditions to prevent action shortcuts. At deployment, our method retains only its lightweight adapters to perform implicit 3D reasoning, discarding the 3D foundation model and the teacher branch used for supervision. Consequently, it operates purely on 2D images without 3D sensors, external models, or explicit text generation while preventing catastrophic forgetting of the pretrained VLM, achieving state-of-the-art performance on LIBERO, LIBERO-PLUS, SimplerEnv, and real-world manipulation tasks.
Jiaxin Shi, Xidong Zhang, Fucai Zhu +3
Shanghai Jiao Tong University · Daimon Robotics · Harbin Institute of Technology +4