MobiGuide: Vocabulary-Guided Online Recovery for Mobile Manipulation

The gist The AnchorVLA framework stabilizes long-horizon mobile manipulation via anchor-guided trajectory generation and test-time refinement, which retains multimodal action generation while reducing inference cost for closed-loop control.

Problem Formulation

Mobile manipulation requires a policy that is simultaneously computationally efficient, capable of modeling multi-modal action distributions, and robust to execution-time drift introduced by action chunking. The core difficulty lies in the strong temporal coupling between locomotion and manipulation, where early motion decisions shape later viewpoints and object accessibility. Recent progress in vision-language-action (VLA) models has been observed primarily in tabletop settings, where feasible motions are concentrated around a narrow set of solutions. Mobile manipulation is fundamentally less structured, meaning the same instruction may admit multiple valid ways of execution depending on the scene and robot state. This makes the action distribution inherently multimodal, and simple teacher-forced (L1) regression often favors an average of several valid behaviors rather than committing to a single executable mode.

AnchorVLA Architecture

AnchorVLA is a diffusion-based VLA policy that combines a lightweight VLA adaptation backbone with an anchored diffusion action head and a test-time self-correction mechanism. The key idea is to leverage training demonstrations to construct a trajectory vocabulary of representative expert motion patterns that serve as structured priors over plausible behaviors. Given the current observation and instruction, AnchorVLA initializes from these precomputed anchors and performs anchored truncated diffusion locally around them, rather than denoising from scratch over the full action space.

Anchored Truncated Diffusion Action Head

The framework introduces a trajectory vocabulary V = M m=1, where each anchor A¯(m) ∈ R H×d represents a structured prior derived from representative expert motion patterns. AnchorVLA simultaneously evaluates all M predefined anchors and performs anchored truncated diffusion, denoising strictly localized around these plausible anchors over a shortened schedule. The noisy actions are generated by perturbing the anchors rather than starting from zero, using the formula A t(τ,m) = sqrt(α τ)barA(m) + sqrt(1-α τ)ϵ where epsilon N(0, I). A dedicated anchor scoring head h(·) predicts a scalar confidence score s(m), which selects the optimal denoised action chunk Aˆt = A hat(m) t = sqrt(A hat(m) t) where mˆ = arg max m s(m).

Test-Time Self-Correction

To mitigate the open-loop drift during the execution of Aˆt, a lightweight residual correction module rϕ(·) is introduced that performs per-step micro-adjustments during rollout. For each execution phase j ∈ 0 to H - 1, the actual applied continuous command is updated dynamically as a t+j = Aˆt+j + Δa t+j where the state-dependent micro-adjustment is predicted as Δa t+j = rψ(o t+j, l, Aˆt+j, j). This module acts as a critical open-loop safeguard that cleanly separates high-level planning from high-frequency reactive corrections.

Experimental Validation

AnchorVLA was validated through both large-scale simulation and real-world deployment, achieving a state-of-the-art 64.0% average success rate on the ManiSkill-HAB mobile manipulation benchmark. Crucially, AnchorVLA remains highly robust to extended execution horizons (H = 5), maintaining a stable 61.5% success rate, effectively rivaling computationally heavy reactive baselines. Ablation studies confirm the necessity of the anchored prior, showing that without it, performance completely fails (0.0% success) under severely truncated schedules. Furthermore, adding the residual module substantially improves task success by 9.3% while maintaining negligible computational cost.

Inference Efficiency and Performance Trade-off

AnchorVLA demonstrates a superior compute-performance trade-off compared to deterministic L1 regression baselines when using longer action chunks (H=5) or more efficiently than reactive updates (H=2). While standard deterministic models suffer from compounding open-loop errors and mode-averaging at longer horizons, AnchorVLA degrades more gradually and remains competitive. By safely leveraging the robust extended execution horizons enabled by AnchorVLA, the system only needs to query visual features intermittently, reducing per-episode compute to just 128.3 TFLOPs compared to 641.4 TFLOPs for H=1. This confirms that the anchored generative prior and integrated scoring head reliably identify optimal multi-modal trajectories without sacrificing execution quality.

Conclusion

AnchorVLA offers a favorable compute-performance trade-off, scaling to longer action chunks without the mode-averaging failures of standard regression. The framework successfully captures the diverse and multimodal nature of expert demonstrations while satisfying real-time constraints for whole-body control. Future work is suggested to incorporate reinforcement learning for stronger test-time self-correction and recovery in out-of-distribution settings. The paper's references include Bjorck, J., Castañeda, F., Cherniadev, N., Da, X., Ding, R., Fan, L., Fang, Y., Fox, D., Hu, F., Huang, S., Jang, J., Jiang, Z. (2025) for GR00T N1: an open foundation model for generalist humanoid robots. The paper also references Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., Fusai, N. (2024) for pi 0: A vision-language-action flow model for general robot control. It also references Chen, S., Liu, J., Qian, S. (2025) for Ac-dit: Adaptive coordination diffusion transformer for mobile manipulation. The paper also references Yuan, X. (2024) for Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. It also references Wu, Z., Zhou, Y., Xu, X. (2026) for A pragmatic vla foundation model. The paper also references MoManipVLA (2025) for Momanipvla: Transferring vision-language-action models for general mobile manipulation. It also references Zhao, Z., Fan, K., Xu, H.Y. (2025) for Anchordp3: 3d affordance guided sparse diffusion policy for robotic manipulation. It also references Xing, Z., Zhang, X., Hu, Y. (2025) for Goalflow: Goal-driven flow matching for multimodal trajectories generation in end-to-end autonomous driving. It also references Xue, H., Ren, J., Chen, W. (2025) for Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation. It also references Yang, J., Huang, I., Vu, B. (2025) for Mobi-pi: Mobilizing your robot learning policy. It also references Yuan, X., Mu, T., Tao, S. (2025) for Policy decorator: Modelagnostic online refinement for large policy model. It also references Li, Z., Wang, S., Lan, S. (2025) for Hydra-next: Robust closedloop driving with open-loop training. It also references Li, Z., Yao, W., Wang, Z. (2025) for Generalized trajectory scoring for end-to-end multimodal planning. It also references Zitkovich, B., Yu, T., Xu, S. (2024) for RT-2: vision-languageaction models transfer web knowledge to robotic control. It also references Zhu, M., Zhu, Y., Li, J. (2025) for Scaling diffusion policy in transformer to 1 billion parameters for robotic manipulation.

1. Bold header: AnchorVLA enables multimodal action modeling for mobile manipulation. This system can generate multiple valid ways of execution by leveraging a trajectory vocabulary derived from expert demonstrations, unlike deterministic regression which often produces smooth but physically inconsistent predictions.

2. Bold header: Truncated diffusion schedule reduces inference cost while preserving multimodality. The paper introduces this mechanism to mitigate the cost of full iterative denoising, allowing for action generation with a reduced diffusion drift and sampling cost compared to standard policies.

3. Bold header: Test-time self-correction mitigates execution-time drift. By introducing a lightweight residual correction module that makes high-frequency, per-step adjustments during rollout, the system can adapt the macro-trajectory to ensure successful end-to-end mobile manipulation despite accumulated errors.

4. Bold header: Enhanced robustness under long execution horizons is achieved through chunking and correction. The paper demonstrates that AnchorVLA maintains a stable 61.5% success rate at longer chunks (H=5) while deterministic models suffer from compounding open-loop errors and mode-averaging.

5. Bold header: Improved performance on complex manipulation tasks via anchor scoring. The system utilizes a dedicated anchor scoring head that evaluates the refined trajectories, predicting a confidence score s(m) to select the optimal denoised action chunk, leading to superior task success rates compared to baselines like AC-DiT.

6. Bold header: Efficient computational scaling for long-horizon planning is achieved through adaptive chunk length. AnchorVLA exhibits a vastly superior compute-performance trade-off over L1 regression, allowing it to remain competitive at H=8 and H=10 while achieving significant inference savings.

A central challenge in mobile manipulation is preserving multiple plausible action modes while remaining reactive during execution. For example, grasping a bottle in a cluttered scene can be achieved through multiple physically feasible base-arm motion plans. Deterministic VLA policies trained with L 1 regression are efficient, but they tend to average across valid modes and may produce physically invalid actions. Diffusion-based generative policies can better represent multimodal action distributions, yet their proposals may be weakly constrained when generated from pure noise, and iterative sampling can increase inference cost. Action chunking helps amortize inference, but it also introduces partially open-loop execution, allowing small mismatches to accumulate into drift. We present MobiGuide, a generative VLA model for efficient mobile manipulation, where action generation is guided by a structured trajectory vocabulary. Constructed from expert demonstrations, this vocabulary encodes plausible whole-body motion patterns and provides structured priors for feasible action generation. Given the current observation and instruction, MobiGuide locally refines these reference trajectories into context-compatible action proposals and selects the most suitable trajectory for execution. This preserves multimodal action candidates while keeping generation constrained around feasible expert-like motion modes. To mitigate chunk-induced drift, we further introduce a lightweight residual correction module that performs high-frequency, per-step adjustments during rollout. Across diverse mobile manipulation tasks, MobiGuide improves success and stability under execution disturbances while maintaining low-latency inference.