CoRL 2026 · Austin, Texas

EvoPlan: Evolutionary Neuro-Symbolic Robot Planning with Spatio-Temporal Guarantees

Bhavya Sai Nukapotula Samin Moosavi Haoze Wang Luke Duncan Diya Shakkottai Varun Murali Srinivas Shakkottai

Texas A&M University

Paper (arXiv) CodeComing soon

TL;DR. An LLM proposes and repairs PDDL plans, programmatic validators decide which plans survive, and a Signal Temporal Logic constraint mined from unannotated demonstrations decides whether each move may be executed. When the robot violates the constraint, it replans.

Real-robot demo

A Clearpath Jackal J100 (ROS 2, Nav2, YOLO-World perception) is asked to reach a backpack whose location is never stated, so the PDDL problem is unsolvable as written. VAL truncates the plan to its executable prefix, which is a survey of the office. When perception detects the backpack, the problem is rewritten and EvoPlan's validated repair is hot-swapped into the running mission without restarting the executor.

Jackal surveying the office — Surveying: the goal object has no stated location.

Jackal driving past cones — The survey continues while YOLO-World searches.

Jackal at the backpack — At the backpack after the perception-grounded replan.

Videos

Open-world search, run 1

The backpack’s location is never given. YOLO-World detects it, the detection is grounded to the conference-door region, and the plan is repaired mid-mission.

Open-world search, run 2

After the backpack is spotted, the repaired plan finishes the survey at the empty wall first, then returns through the conference door to it.

STL shield on the real robot

A person blocks the doorway. Every route past them violates the clearance constraint, so the move is vetoed and the planner switches to a reachable goal.

Left: phone camera; right: RViz costmap. Videos are silent and sped up where noted; use Enlarge to watch at full size.

Abstract

LLM-based robot planners are fluent but cannot guarantee that their plans are executable or safe. Classical PDDL planners can guarantee these properties, but only after the problem is fully specified, and they make poor use of an LLM's ability to read context and repair plans. This paper presents a neuro-symbolic framework with three parts.

First, an offline procedure that mines a single global Signal Temporal Logic (STL) constraint on mobility from demonstration data. The procedure recovers codified rules (e.g., stopping at red lights, mined from nuPlan driving logs) or population preferences (e.g., social-navigation comfort, mined from SCAND teleoperation), depending on what the demonstrations encode. Because the demonstrations are a one-class signal, we generate the missing negatives with counterfactual perturbations and an LLM violation generator and then fit the constraint by evolutionary search. We use the mined constraint to shield a vision-language driving policy on Bench2Drive and two discrete-action navigation policies on HA-VLN-CE.

Second, an evolutionary PDDL planner: an LLM proposes and repairs plans, programmatic validators decide which ones survive, and the validated portion of the plan grows over iterations. We test the planner on the open-world ALFWorld Text benchmark, where it beats strong baselines and stays robust when the goal vocabulary does not match the action-model vocabulary.

Third, a constrained execution loop: the planner's plan is compiled into waypoints, the waypoints are checked against the mined constraint, and the planner re-plans on a violation.

Method

EvoPlan system overview: offline constraint extraction (top) and evolutionary PDDL solving with constrained deployment (bottom) — **System overview.** *Offline* (top), demonstrations are reduced to kinematic and social signals, counterfactual negatives turn the one-class data into contrastive sets, and evolutionary STL fitting produces one global mobility constraint. *Online* (bottom), an evolutionary PDDL solver proposes plans that PDDL validation must accept; each action is expanded into waypoints, checked against the STL constraint, and executed under monitoring with replanning.

1 · Offline

Mine an STL constraint

Demonstrations contain only acceptable behaviour. Counterfactual perturbations supply the missing negatives, and LLM-guided evolutionary search fits one global STL formula over speed, clearance and social signals.

2 · Planning

Evolve PDDL plans

An LLM proposes and repairs plans. A validator cascade (syntax, VAL simulation, goal reachability) decides which ones survive, so the LLM suggests plans but never decides which are accepted.

3 · Execution

Check, execute, replan

Each action becomes a waypoint sequence that is checked against the mined constraint before execution. A violation commits the verified prefix and triggers a replan from the updated state.

Results

98.5%open-world ALFWorld Text, aligned and misaligned vocabulary

−71%collisions on Bench2Drive with the mined rule shield

−82% / −93%avoidable contacts on HA-VLN-CE (PointNav / MLP)

100%Blocksworld Misalignment, where Fast Downward scores 0%

Evolutionary PDDL planner on open-world ALFWorld Text

Success rate on the 134 unseen tasks. In the misaligned variant, the goal vocabulary (hot / clean / cool) does not match the action vocabulary (heated / washed / chilled). Baseline numbers are reproduced from NL-PDDL (Liu et al., ICLR 2026).

Method	Aligned SR %	Misaligned SR %
GPT-4o (direct)	21	17
ReAct (with examples)	81	79
Reflexion-10	91	—
NL-PDDL	94	91
EvoPlan (Qwen3-32B)	98.5	98.5

Verified-prefix growth, open-world — Open-world

Verified-prefix growth, closed-world — Closed-world

Closed-world population state — Closed-world population

The validated fraction of the plan grows monotonically over iterations on hard tasks, although no prefix is explicitly frozen.

Mined constraints shield downstream policies

The same mining procedure recovers rules from nuPlan driving logs and preferences from SCAND teleoperation. Neither shield is trained on the target benchmark.

Bench2Drive	DS ↑	Coll. ↓	Infr. ↓
Unshielded	62.96	14	20
Shielded, α=0	65.61	4	7
Shielded, α=0.30	66.20	4	6

HA-VLN-CE	Shield	SR % ↑	CT % ↓
PointNav	off	58.7	38.1
PointNav	on	57.6	18.5
MLP	off	32.1	30.1
MLP	on	32.7	5.2

Per-episode safety distributions with and without the shield — Per-episode safety distributions on HA-VLN-CE `val_unseen` (N = 1839 per violin). The shield removes the heavy unsafe tail on both architectures rather than only shifting the mean.

Constrained execution loop in Gazebo

Shield vetoes a step near a pedestrian — Shield in action

PDDL move actions are grounded into routes by Nav2 and checked against the mined constraint before execution. When a pedestrian enters the learned clearance floor, the step is vetoed and the planner replans.

Simulation run A

The shield flags a path too close to a pedestrian, the move r5 → r1 becomes undrivable, and EvoPlan reorders the tour to visit r2 first. Zero collisions.

Simulation run B

Same mission, different pedestrian traffic. A live panel shows every STL check, the clearance margin over time, and the mission log. Never closer than 1.08 m to a person.

Left: Gazebo; right: Nav2 costmap (run A) or live STL panel (run B). Footage is sped up where noted.

BibTeX

@inproceedings{nukapotula2026evoplan,
  title         = {EvoPlan: Evolutionary Neuro-Symbolic Robot Planning
                   with Spatio-Temporal Guarantees},
  author        = {Nukapotula, Bhavya Sai and Moosavi, Samin and Wang, Haoze and
                   Duncan, Luke and Shakkottai, Diya and Murali, Varun and
                   Shakkottai, Srinivas},
  booktitle     = {Proceedings of the 10th Conference on Robot Learning (CoRL)},
  series        = {Proceedings of Machine Learning Research},
  year          = {2026},
  eprint        = {2607.06724},
  archivePrefix = {arXiv}
}