Skip to main content
Back to timeline
arXivSource publication:

REINS moves the agent harness into driving planning: LLM call rate falls from 70% to 12.81% with no weight updates

Synopsis

REINS ports the agent-harness pattern familiar from software engineering into real-time trajectory planning: an LLM's driving suggestion is grounded to a learned variable-level dynamics model, checked by calibrated predicates and multi-step forward rollout, then written into an indexed rule memory. Over 1000 episodes on the highway-env Highway scenario, the LLM call rate fell from 70.0% to 12.81%, compliance rose from 0.9885 to 0.9956, and collision rate fell from 7.05% to 5.67%, with no LLM weight update.

Source-provided article image: REINS: A Self-Evolving Agent Harness for Real-Time Trajectory Planning
arXiv

Interpretation

REINS lets the LLM appear only on memory misses, with verified rule memory answering the rest, keeping second-scale language-model queries off the millisecond control loop. Earlier LLM trajectory planners either kept the LLM in the decision loop, where a single query far exceeds the millisecond control budget, or extracted rules offline and discarded the LLM, leaving the runtime planner unable to evolve with deployment. Over 1000 episodes on the highway-env Highway scenario, the LLM call rate fell monotonically from 70.0% at episode 250 to 12.81% at episode 1000, while compliance rose from 0.9885 to 0.9956 and collision rate fell from 7.05% to 5.67%.

The runtime has three couplings: grounding LLM output to physical variables, verifying maneuvers by forward rollout, and consolidating verified maneuvers into retrievable rules. Grounding relies on a dynamics model learned over a factored variable space, where the dependency graph is set jointly by LLM semantic priors and group-weighted ridge regression, and each variable is predicted by its own Transformer. Verification compiles LLM proposals into predicates with data-calibrated thresholds, releasing only commitments whose predicates all hold along the predicted trajectory; of roughly 39,000 proposals in the 1000-episode deployment, 2.1% failed schema validation and 18.0% were rejected by rollout verification, and the rest deduplicated to 57 new rules.

On a memory hit, parameterized graph queries retrieve a few rules for the active variables, and the planner scores actions, rolls out candidates over a fixed horizon, and executes the first step. Rules live in a Neo4j graph database, with rule nodes linked to variable nodes by TRIGGERS and AFFECTS edges, so retrieval touches only 2 to 3 rules per step. On simulated and real-world benchmarks, planning latency is 1.9 to 4.6 ms per step, compliance exceeds 98%, and collision rate is 49% to 56% below the strongest baseline, Plan-R1.

Perspective

Evaluation sits at the tactical decision layer: perception is assumed to supply clean factored variables, highD and rounD run in closed-loop replay, and occlusion, tracking and map errors are absent. The authors state these are harness results at this layer, not a deployable autonomous-driving stack, and that episode-level collision rates of 5.7% to 9.5% are not deployment grade; the supported claim is relative, 49% to 56% below the best learned baseline.

The learned dynamics are prior-constrained but not causally identified, and verification quality depends on dynamics fidelity, with constant-velocity propagation raising collision rate by 48% to 49%; 3 of the 57 acquired rules were over-specific in audit. The monotone miss-rate property assumes a stationary scene distribution, and a memory compiled only from highD and rounD starts at an 86.2% call rate on highway-env, indicating limited zero-shot cross-domain reuse.