Since current Vision-Language-Action (VLA) systems suffer from limited spatial perception and the absence of memory throughout manipulation, we investigate visual anchors as a means to enhance spatial and temporal reasoning within VLA policies for robotic manipulation. Conventional VLAs generate actions by conditioning on a single current frame together with a language instruction. However, since the frame is encoded as a 2D image, it does not contain detailed spatial information, and the VLA similarly lacks any means to incorporate past context. As a result, it frequently forgets objects under occlusion and becomes spatially disoriented during the manipulation process. Thus, we propose AnchorVLA4D, a simple spatial-temporal VLA that augments the visual input with an anchor image to preserve the initial scene context throughout execution, and adds a lightweight spatial encoder that jointly processes the anchor and current frames to expose geometric relationships within an episode. Built on a Qwen2.5-VL backbone with a diffusion-based action head, AnchorVLA4D requires no additional sensing modalities (e.g., depth or point clouds) and introduces negligible inference overhead. Combining anchoring with a frozen pretrained spatial encoder yields further gains, realizing a 13.6% improvement on the Simpler WidowX benchmark and confirming the approach on real-world tasks, where it achieved an average success rate of 80%.
Figures & tables
Fig. 1: Limits of depending on a single frame in conventional VLAs. The top row illustrates forgetting caused by occlusion, while the bottom row demonstrates spatial disorientation.
Fig. 2: Overall Architecture
Method Type
Model Size
Models
Spoon on Towel
Carrot on Plate
Stack Cube
Eggplant in Basket
Average
Standard
7B
OpenVLA [ 1 ]
4.2%
0.0%
0.0%
12.5%
4.2%
7B + 300M
CogACT-Base [ 3 ]
71.7%
50.8%
15.0%
67.5%
51.3%
3B + 300M
π0 -Uniform ⋆ [ 2 ]
63.3%
58.8%
21.3%
79.2%
55.7%
Spatial
3.5B
SpatialVLA [ 13 ]
16.7%
25.0%
29.2%
100%
42.7%
3.5B
SoFar [ 25 ]
58.3%
66.7%
70.8%
37.5%
58.3%
13B
FSD [ 23 ]
41.7%
50.0%
33.3%
37.5%
40.6%
TABLE I: Simulation Success Rate
Fig. 3: More precise retries using an anchor
Fig. 4: Tasks in Real-World Environments
Fig. 5: Success rates for three different tasks
Models
Anchor
SE
SR
Performance+-
VanillaVLA †
w/o
w/o
51.0%
-
AnchorVLA
w/
w/o
60.4%
† + 9.4%
AnchorVLA4D ∗
w/
w/
64.6%
† + 13.6%
AnchorVLA4D
w/
unfreeze
59.4%
∗ - 6.2%
AnchorVLA4D
w/
remove
58.3%
∗ - 6.3%
TABLE II: Performance Breakdown
Anchor Format
SE Injection
SR
I0
-
55.2%
Ii−3…i−1
-
46.9%
Ii−60,i−40,i−20
-
21.9%
I0
Injection before Decoder
0%
I0
Injection in Action Head
46.9%
I0
Concatenation
64.6%
TABLE III: Integration Ablation
Anchor
Overall Retries
in Case of Success
in Case of Failure
w/o
1.54
1.43
1.66
w/
1.40
1.21
1.68
TABLE IV: Number of Retries
Proprioceptive States
Simulation(10k)
Realworld Open Drawer(5k)
w/o
40.6%
60%
w/
51.0%
50%
TABLE V: Comparison With or Without Proprioceptive States
Reactive vision-language-action (VLA) policies suffer from task-state aliasing in long-horizon manipulation, where identical multimodal inputs call for distinct, context-dependent actions. Given that pretrained VLAs already possess rich control primitives to express diverse behaviors, we hypothesize that the execution bottleneck lies not in policy capacity, but in input ambiguity. In this paper, we propose TaskAnchor, a lightweight adapter that grounds task state by injecting execution context into the VLA's native input space. During post-training, TaskAnchor learns to represent the semantic execution stage as a milestone-supervised coordinate prepended to the language instruction, while incorporating fine-grained historical evidence via a residual update to the current visual tokens. This formulation avoids generating complex subtask instructions and leaves the backbone architecture unchanged. Across long-horizon benchmarks, TaskAnchor delivers substantial gains, achieving approximately 6 times the average success rate of the pi0.5 and X-VLA baselines on RMBench and more than doubling the task success rate of pi0.5 on RoboMemArena. Real-robot experiments further validate reliable multi-stage execution, with the same policy adapting its subsequent behaviors using earlier human interactions as in-context cues. Our project website is available at https://taskanchor.netlify.app/.
Hengyan Liu, Wenlve Zhou, Bo Yue +7
The Chinese University of Hong Kong, Shenzhen · Foshan University · DexForce +1
Vision-language-action (VLA) models increasingly condition robot policies on history, depth, or 4D features to resolve ambiguity in long-horizon manipulation. However, more spatiotemporal evidence is not necessarily better: when the injected evidence is not motion-consistent, it can introduce geometric drift, fragmented temporal cues, and unstable action generation. This raises a simple question: should a VLA remember past frames, or remember the motion that connects them? We introduce MotionVLA, a motion-history interface that converts a short past-only video window into compact, time-continuous trajectory-field tokens. Instead of treating history as a sparse set of ndependently lifted frames, MotionVLA represents recent observations as physically coherent motion evidence. Current visual tokens query this history to retrieve task-relevant motion information, which is then recoupled into the VLA stream under trajectory-grounded supervision. Experiments across simulation benchmarks and preliminary real-robot rollouts show that MotionVLA improves long-horizon manipulation while producing smoother and more direct executions. These results suggest that effective VLA memory is not just about providing more 4D context, but about exposing motion-consistent evidence that is usable for control.
Shanglin Yuan, Weiheng Zhao, Xianda Guo +4
Huazhong University of Science and Technology · D-Robotics · Wuhan University
Robotic manipulation in open-world environments requires reasoning across semantics, geometry, and long-horizon action dynamics. Existing hierarchical Vision-Language-Action (VLA) frameworks typically use 2D representations to connect high-level reasoning with low-level control, but lack depth awareness and temporal consistency, limiting robustness in complex 3D scenes. We propose ST-VLA, a hierarchical VLA framework using a unified 3D-4D representation to bridge perception and action. ST-VLA converts 2D guidance into 3D trajectories and generates smooth spatial masks that capture 4D spatio-temporal context, providing a stable interface between semantic reasoning and continuous control. To enable effective learning of such representations, we introduce ST-Human, a large-scale human manipulation dataset with 14 tasks and 300k episodes, annotated with 2D, 3D, and 4D supervision via a semi-automated pipeline. Using ST-Human, we train ST-VLM, a spatio-temporal vision-language model that generates spatially grounded and temporally coherent 3D representations to guide policy execution. The smooth spatial masks focus on task-relevant geometry and stabilize latent representations, enabling online replanning and long-horizon reasoning. Experiments on RLBench and real-world manipulation tasks show that \method significantly outperforms state-of-the-art baselines, improving zero-shot success rates by 44.6% and 30.3%. These results demonstrate that offloading spatio-temporal reasoning to VLMs with unified 3D-4D representations substantially improves robustness and generalization for open-world robotic manipulation. Project website: https://oucx117.github.io/ST-VLA/.
You Wu, Zixuan Chen, Cunxu Ou +9
School of Computer Science, Nanjing University · School of Intelligence Science and Technology, Nanjing University · School of Electronic Science and Engineering, Nanjing University +2