AnchorVLA: Bridging Discrete Decisions and Continuous Trajectories for Vision-Language-Action Planning

Abstract: Autonomous driving planning requires translating navigation intent, traffic rules, dynamic interactions, and language instructions into executable continuous trajectories. Vision-Language-Action models have been introduced into driving planning to improve long-tail generalization, commonsense reasoning, high-level semantic understanding, and explainability. However, existing VLA planners mainly follow planning-head-based trajectory prediction or full-trajectory autoregressive generation. The former only weakly constrains continuous trajectory generation with VLA reasoning, while the latter relies on long sequences of low-information-density coordinate tokens, making semantic-action alignment difficult and leading to discretization errors and inefficient inference. To address these limitations, we propose AnchorVLA, a hierarchical decision-anchored VLA planning framework that uses trajectory-pattern anchors as an explicit interface between high-level VLA reasoning and continuous trajectory execution. Specifically, Decision-as-Anchor Representation represents behavior-level driving decisions with anchor tokens, each encoding an entire local motion pattern rather than a single coordinate point. Decision-Anchored Residual Flow then generates fine-grained continuous trajectories in the selected anchor-defined residual space, capturing multi-modal execution refinements after high-level decision making. By reasoning over compact and semantically meaningful anchors instead of autoregressively generating waypoint sequences, AnchorVLA preserves LLM-based decision making while improving inference efficiency, semantic-action alignment, and continuous generation flexibility. Experiments on the Bench2Drive closed-loop benchmark show that AnchorVLA achieves a state-of-the-art Success Rate of 77.28 and a competitive Driving Score of 89.92.

• The paper introduces a hierarchical framework that uses trajectory-pattern anchors to bridge discrete decision-making with continuous trajectory execution.
• It combines the Decision-as-Anchor Representation (DAAR) and Decision-Anchored Residual Flow (DARF) to reduce action space complexity and enhance semantic-action alignment.
• Experimental results show improved success rate and reduced latency, demonstrating robust behavior-level decision making on autonomous driving benchmarks.

AnchorVLA: Hierarchical Decision-Anchored VLA Planning for Autonomous Driving

Motivation and Paradigm Analysis

The AnchorVLA framework is designed to resolve fundamental limitations in current Vision-Language-Action (VLA) planning paradigms for autonomous driving. Conventional end-to-end planners unify perception, prediction, mapping, and planning, but suffer from limited semantic decision capability. Planning-head-based VLA planners leverage pre-trained LLM/VLM features for trajectory prediction, while autoregressive planners generate coordinates or waypoints as discrete tokens. Both approaches face an abstraction gap: planning-head-based methods weakly constrain trajectories with high-level VLA reasoning, and autoregressive methods suffer from low semantic density tokens and inefficient inference due to long token sequences.

AnchorVLA addresses this by introducing trajectory-pattern anchors as explicit decision interfaces bridging high-level VLA reasoning and continuous trajectory execution. (Figure 1)

Figure 1: Comparison of autonomous driving planning paradigms, illustrating AnchorVLA's explicit decision interface via trajectory-pattern anchors.

Methodology

Decision-as-Anchor Representation (DAAR)

DAAR models high-level driving decisions as selection from a codebook of trajectory-pattern anchors, each representing a distinct local motion pattern (e.g., lane keeping, braking, overtaking). Rather than tokenizing individual coordinates, DAAR clusters trajectories to construct a compact, behavior-level anchor space. This reduces the action space complexity and strengthens semantic-action alignment.

Both query-based and autoregressive anchor-decision modeling methods are instantiated. Autoregressive modeling incorporates anchor tokens directly into the LLM vocabulary, enabling anchor prediction via next-token prediction, resulting in tighter coupling between language reasoning and discrete action abstraction.

Figure 2: Overview of AnchorVLA, wherein high-level VLA decisions are expressed as trajectory-pattern anchors and continuous trajectories are generated in anchor-defined residual spaces.

Decision-Anchored Residual Flow (DARF)

DARF generates fine-grained, continuous trajectories within the residual coordinate system defined by the selected anchor. Flow matching is performed in these local subspaces, enabling multi-modal residual refinement consistent with the anchor decision. The output trajectory is thus the sum of the anchor and the learned residual, effectively mirroring decision-then-execution behavior in human driving.

DARF predicts candidate trajectories for top-M anchors and selects the final output using a context-conditioned confidence score. This structure preserves high-level decision multimodality up to final execution.

Figure 3: Illustration of DARF, demonstrating decomposition of the global trajectory space into anchor-defined residual subspaces for fine-grained generation.

Experimental Results

On the Bench2Drive closed-loop benchmark, AnchorVLA achieves a Success Rate (SR) of 77.28% and a Driving Score (DS) of 89.92, establishing strong performance especially in task completion reliability. Compared to LinkVLA and BridgeDrive, AnchorVLA improves SR while maintaining competitive DS. Efficiency and comfort scores reveal a trade-off: residual refinement increases generative flexibility but also introduces potential smoothness degradation.

Multimodal ability evaluation shows superiority across Overtake, Brake, and overall mean scores, indicating robust behavior-level decision alignment in interactive scenarios.

Ablation studies highlight that confining autoregressive reasoning to anchor tokens achieves substantial latency reduction (from 361 ms to 64 ms) and higher SR, versus full-sequence autoregression. DARF outperforms deterministic regression and full-flow matching, demonstrating the effectiveness of anchor-conditioned residual modeling.

Qualitative Visualizations

Visualizations illustrate the separation between behavior-level anchors, chain-of-thought (CoT) VLA reasoning, and continuous trajectory execution. Anchor decisions consistently reflect high-level maneuvers (e.g., braking, turning), and DARF refines these within well-defined residual spaces, ensuring geometric precision without semantic drift.

Figure 4: Qualitative visualization of AnchorVLA, showing CoT-based decision, anchor prediction, and final trajectory refinement.

Failure modes are dominated by error propagation from inaccurate anchor selection, such as off-route or language-inconsistent anchors, and noisy residual estimation. These underscore the need for enhanced reflection and anchor selection mechanisms to safeguard semantic consistency during residual refinement.

Figure 5: Failure case analysis of AnchorVLA, highlighting modes including off-route, language mismatch, and trajectory deviation.

Theoretical and Practical Implications

AnchorVLA's hierarchical decomposition—explicit anchor selection followed by continuous residual generation—offers a principled bridge between discrete language-driven planning and flexible geometric execution. This structured interface augments explainability, semantic-action alignment, and inference efficiency. Practically, it enables end-to-end planners to incorporate rich language/contextual reasoning without sacrificing real-time operational constraints or closed-loop reliability.

The anchor-coded interface creates new potential for hierarchical reinforcement and feedback, allowing reflection mechanisms and context-adaptive revision for robust planning in ambiguous, long-tail scenarios.

Future Directions

Future developments may include:

• Enhanced multimodal reasoning for anchor selection, integrating scene context, historical behaviors, and linguistic nuance.
• Reflection or revision mechanisms for anchor prediction to mitigate error propagation.
• Incorporation of smoothness/comfort objectives in DARF for improved trajectory quality.
• Extension of anchor-based interfaces to complex, multi-agent or adversarial planning settings, leveraging anchor prediction for coordinated behaviors.

Conclusion

AnchorVLA introduces a hierarchical decision-anchored VLA planning framework that formulates trajectory-pattern anchors as explicit decision interfaces, bridging high-level semantic reasoning and continuous trajectory execution. The DAAR + DARF modularity demonstrates improved reliability, semantic-action alignment, and inference efficiency, validated on challenging benchmarks and interactive driving scenarios. The anchor-based decomposition presents promising directions for scalable, explainable, and robust vision-language-driven autonomous agents.

(2607.03182)