REINS: A Self-Evolving Agent Harness for Real-Time Trajectory Planning
REINS: A Self-Evolving Agent Harness for Real-Time Trajectory Planning
This research introduces REINS, a runtime harness designed for real-time trajectory planning that wraps an LLM's reasoning loop with tool dispatch, verification, and persistent memory. The system grounds LLM outputs to a learned variable-level dynamics model, verifies them via calibrated predicates and forward rollout, and consolidates results into an indexed rule memory serving future scenes in O(log n) time. By closing a self-evolution loop where familiar scenes resolve via memory lookup and novel scenes invoke the LLM to enrich memory, REINS achieves over 98% compliance, nearly 50% collision reduction, and 1.9-4.6 ms latency on simulated and real-world benchmarks without modifying LLM weights.
The paper proposes REINS, a runtime harness that grounds LLM outputs to learned dynamics models and consolidates verified trajectories into an indexed rule memory to meet real-time constraints while self-evolving.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
latency
1.9–4.6
View original evidence
vel scenes invoke the LLM, whose verified outputs enrich the memory. Without modifying LLM weights, REINS attains 98%+ compliance, nearly 50% collision reduction, and 1.9-4.6 ms latency on simulated and real-world benchmarks at the tactical decision layer, with LLM invocation rate monotonically declining as memory matures. Code: https://github.com/ZhihongCui/REINS
Open source S5
- code: Available Artifact link Open source S5
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
vel scenes invoke the LLM, whose verified outputs enrich the memory. Without modifying LLM weights, REINS attains 98%+ compliance, nearly 50% collision reduction, and 1.9-4.6 ms latency on simulated and real-world benchmarks at the tactical decision layer, with LLM invocation rate monotonically declining as memory matures. Code: https://github.com/ZhihongCui/REINS
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org