Cognitive-Horizon Action Chunking: Memory-Guided Adaptive Reasoning for Embodied Navigation

Weile Gong, Zijian Lu, Jiahao Zheng, Bo Xu, Yiping Zuo, Jiajia Guo, Xin He, Weijun Wang, Weibei FanarXiv 2026 · arXiv

TL;DR

Cognitive-Horizon Action Chunking (CHAC) is a training-free framework that makes robot navigation significantly more efficient by adaptively reusing predictions from large vision-language models (VLMs). By intelligently deciding when to refresh plans based on visual context and using past movement history to fine-tune final stopping positions, CHAC reduces expensive model calls by up to 65% with minimal impact on success rates.

What Problem Does This Paper Address?

State-of-the-art embodied navigation relies on massive multimodal foundation models to interpret scenes and generate actions. However, these models are computationally "heavy." Running inference at every single step (native execution) is slow, power-hungry, and often redundant.

While "action chunking"—predicting a sequence of future waypoints at once—can save costs, fixed-length chunks are brittle. If a robot turns a corner and sees a new room, or gets stuck against an obstacle, it shouldn't blindly follow an old plan. The challenge is finding the "cognitive horizon": exactly how many steps a robot should take from a cached plan before it needs to think again. Furthermore, robots often fail tasks by stopping just outside the target zone; CHAC addresses this "last-mile" problem without needing extra expensive model calls.

Methods

CHAC is a training-free wrapper for frozen navigation policies. It introduces two main components:

  1. Memory-Guided Action Reuse: Instead of a fixed chunk size, CHAC determines an online "cognitive horizon" using two boundaries:
  2. Commitment Boundary: Monitors internal cues (like predicted arrival scores) and external feedback (like whether the controller can actually move forward). If the robot is stuck or thinks it has arrived, it triggers a refresh.
  3. Visual Boundary: Compares the current view to the visual context available when the plan was first made. If the robot turns and reveals a "directional mismatch"—meaning the scene looks significantly different from what was previously seen—it forces a new model call.
  4. History-Guided Terminal Correction: When the VLM issues a "STOP" command, CHAC doesn't just stop. It looks at its local visitation history (the "memory" of where it just was). It calculates a local centroid of nearby visited positions and performs one final, bounded movement toward that consensus. This geometric correction helps the robot settle into the true goal region if it overshot it.
Figure 1. Overview of CHAC. A frozen navigation policy predicts ordered waypoints from current views, visual history, and the task instruction. During execution, retained visual context and commitment checks guide when to refresh the cached prediction. At native STOP, nearby visited positions guide a bounded terminal correction without another model call. The lower scene shows how executed motion becomes visitation history. Red dots mark visited positions, and the blue disk denotes the local-history neighborhood.

Datasets

The authors evaluated CHAC across simulation benchmarks and physical hardware:

  • HM3D-OVON: A large-scale dataset for Open-Vocabulary Object Goal Navigation (3,000 test episodes).
  • R2R-CE: A benchmark for Vision-and-Language Navigation in continuous environments following natural language instructions (1,839 test episodes).
  • Physical Robot Trials: Real-world evaluation using a TurboPi mobile robot across four indoor tasks (40 total runs).

Key Insights and Ideas

  • The "Cost Reversal" Phenomenon: Counter-intuitively, adding more triggers for model refreshes (like the visual boundary) can actually decrease the total number of model calls over an entire episode. By refreshing when the scene changes, the robot avoids wandering aimlessly on an outdated plan, leading to shorter paths and fewer total calls.
  • History as a "Free" Geometric Anchor: The robot's own path is a rich, zero-cost data source. Using the local centroid of visited positions provides a surprisingly effective way to fix terminal positioning errors without needing any additional VLM inference.
  • Training-Free Adaptability: CHAC works with frozen models. This is a major advantage for deployment, as it improves efficiency and success without the need for expensive re-training or fine-tuning of massive foundation models.
Visual boundary trigger example where a change in heading reveals new information, prompting a refresh.

Key Results

CHAC demonstrates a superior trade-off between navigation quality and computational cost compared to existing adaptive reasoning methods:

  • Efficiency Gains: Reduced VLM calls by 33.4% on HM3D-OVON and 64.7% on R2R-CE.
  • Performance Preservation: On HM3D, the success rate only dropped by 0.63 percentage points compared to the expensive "native" execution (31.53% vs 32.17%).
  • Physical Success: In real-world trials, CHAC achieved 18 successes compared to 14 for native execution, while slashing GPU inference time by 49.7%.
  • Superior Metrics: CHAC achieved the highest Success Rate (SR) and Success weighted by Path Length (SPL) per model call among all evaluated action-chunking alternatives.
Figure 2. Global cost changes and boundary interaction. (a) Full CHAC versus w/o Commitment on all 1,000 pairs. Both retain Visual and Terminal. Peach indicates lower cost, rose indicates higher cost, coral indicates mixed changes, and gray indicates unchanged costs. (b) Adding Visual with Commitment off or on, and the difference between these effects. All four configurations disable Terminal. Error bars are paired-bootstrap 95% intervals.

Limitations and Future Work

  • Task-Dependent Trade-offs: While success rates remained high on HM3D, the drop was more significant on the R2R-CE benchmark (approx. 10 points), suggesting the "cognitive horizon" needs to be tuned for different types of navigation tasks.
  • Increased Environment Steps: While model calls go down, the total number of physical actions (steps) can sometimes increase, as the robot may take slightly less efficient paths when reusing cached waypoints.
  • Odometry Reliance: The terminal correction and visual boundaries rely on reasonably accurate motion estimates. Significant odometry drift in the real world could degrade performance.
  • Future Directions: The authors suggest exploring broader deployment with more robust measured odometry and investigating how these adaptive horizons could be integrated into end-to-end training in the future.

Summary updated

Abstract

Multimodal foundation models are advancing embodied navigation, but frequent inference incurs substantial computational cost. We present Cognitive-Horizon Action Chunking (CHAC), a training-free framework for memory-guided adaptive reasoning with a frozen navigation policy. Retained visual context, completion cues, and controller feedback determine the cognitive horizon, or how many cached waypoints the robot executes before refreshing its prediction. At the policy's stop request, local visitation history guides a bounded correction with at most one extra movement and no additional model calls. CHAC reduces vision-language model calls by 33.4% on HM3D-OVON and 64.7% on R2R-CE, with success-rate decreases of 0.63 and 10.82 percentage points, respectively. Success rate and success weighted by path length improve per model call on both benchmarks. Across four physical navigation tasks with 40 runs per method, CHAC achieves 18 successes versus 14 for native execution, reducing model calls by 54.9% and GPU inference time by 49.7%.
BibTeX
@misc{gong2026cognitive,
  author = {Weile Gong and Zijian Lu and Jiahao Zheng and Bo Xu and Yiping Zuo and Jiajia Guo and Xin He and Weijun Wang and Weibei Fan},
  title = {{Cognitive-Horizon Action Chunking: Memory-Guided Adaptive Reasoning for Embodied Navigation}},
  year = {2026},
  eprint = {2611.25587},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.25587},
}
Figure