EvoPlan: Evolutionary Neuro-Symbolic Robot Planning with Spatio-Temporal Guarantees
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "EvoPlan: Evolutionary Neuro-Symbolic Robot Planning with Spatio-Temporal Guarantees".
Dev: Here's the gist of the paper. LLM-based robot planners are fluent but cannot guarantee that their plans are executable or safe because classical PDDL planners can guarantee these properties only after the problem is fully specified and they make poor use of an LLM’s ability to read context and repair plans.
Rosa: This paper presents a neuro-symbolic framework with three parts.
Dev: System Overview The framework involves an offline procedure that mines a single global Signal Temporal Logic (STL) constraint on mobility from demonstration data and an evolutionary PDDL planner where an LLM proposes and repairs plans against a fixed PDDL problem P and domain D.
Rosa: Finally, a constrained execution loop checks the plan's waypoints against the mined constraint Φmob before execution.
Paper summary: Rosa: So, we're looking at this paper called "EvoPlan: Evolutionary Neuro-Symbolic Robot Planning with Spatio-Temporal Guarantees." Basically, the core idea is that big language models are fluent but they can't actually guarantee a plan will work or be safe when you put them on a robot.
Dev: Right, so classical planners like PDDL can do that, but they need you to fully specify everything upfront, which is tough in the real world. This paper claims this framework fixes that by using an LLM to help propose and repair plans against a fixed problem setup while adding formal safety checks.
Taro: What matters here for me is how it handles when the robot runs into something unexpected; it suggests a way for the system to recover gracefully instead of just failing completely.
Rosa: It claims this framework works in three main parts: first, an offline procedure that mines a single global signal temporal logic constraint on mobility from demonstration data; second, an evolutionary PDDL planner where an LLM proposes and repairs plans against a fixed problem setup; and finally, a constrained execution loop that checks the plan's waypoints against that mined constraint before it executes.
Dev: The key part I see there is this offline constraint extraction procedure. It mines one global Signal Temporal Logic, or STL, constraint on mobility directly from demonstration data to get rules like stopping at red lights or social navigation comfort preferences.
Rosa: That procedure is designed to synthesize missing negative classes using counterfactual perturbations and an LLM violation generator, which turns the process of finding those constraints into a supervised classification problem. It’s trying to codify how the robot must move at all times, applying that constraint uniformly to every mobility action any plan invokes.
Taro: So, if we think about what this means for real robots outside a lab setting, it suggests we can learn fundamental movement rules from just watching someone do something and turn that into a formal safety contract.
Dev: Exactly. And then online, the evolutionary PDDL solver uses an LLM to propose and repair plans against a fixed PDDL problem and domain. A programmatic validator cascade scores each candidate, and elitism on that feedback grows the validated prefix of the plan monotonically over iterations.
Rosa: And if that constraint check fails during execution, the verified prefix gets committed, and then the planner is re-invoked to produce a fresh suffix-replacing plan. This creates a closed loop where safety verification drives replanning when things go wrong.
Paper summary: Taro: I like that part about replanning; it means even if the initial suggestion from the LLM is bad, we have a mechanism to fix it quickly based on those hard constraints. But what about how robust this whole thing is?
Dev: The paper tests this system on open-world ALFWorld Text where it actually outperforms strong baselines and stays robust even when the goal vocabulary doesn't match the action model vocabulary. They also evaluated the mined constraint as a shield over a Qwen2 point 5-VL driver on Bench2Drive and two discrete-action navigation policies on HA-VLNCE <ref:2607.06724#pg1>.
Rosa: And they showed that in crowded operation, this constrained execution loop demonstrates graceful degradation under crowded conditions, trading clearance margin for an isolated failure rather than collapsing wholesale. That's a big deal for deployment outside controlled settings.
Taro: Graceful degradation is important; it means the robot doesn't just crash when things get messy; it handles the mess by sticking to the learned constraints as much as possible.
Dev: The paper also mentions that all LLM calls use a locally-hosted open-weight model, specifically Qwen3-32B in their experiments, so you don't need some expensive cloud connection for this pipeline to run on a robot <ref:2607.06724#pg1,all LLM calls use a locally-hosted open-weight model>.
Rosa: So, the system is designed to be deployable on hardware right away without needing constant internet access or heavy cloud resources. That makes it much more practical for real-time robotic applications.
Taro: But I wonder about the training signal itself; they mention how they build it by mining a single global STL constraint from demonstrations, which seems like a very specific starting point for the whole system.
Dev: They say that constraint extraction procedure synthesizes missing negatives with counterfactual perturbations and an LLM violation generator to turn the constraint mining into a supervised classification problem. It recovers codified rules or population preferences depending on what the demonstrations encode, which is flexible.
Rosa: So it's not just one hard rule; it can learn different types of constraints based on whether the robot is learning stopping at red lights or maybe learning social navigation comfort from teleoperation data.
Taro: It’s interesting how they say this procedure recovers codified rules or population preferences depending on what the demonstrations encode, which means it adapts its type of constraint mining to the specific behavior it's looking at.
Dev: And then in the evolutionary PDDL solving part, they use an LLM not as a standalone planner but as a proposal mechanism inside this machine-checkable loop, which is how they manage the safety concerns we talked about earlier.
Paper summary: Rosa: It’s interesting because they are treating neural models as generators of candidate plans rather than sources of unchecked authority, which seems like a responsible way to build these things.
Taro: That sounds like a key distinction; you’re not letting the LLM just decide what to do, but you're putting it inside a validation structure. It’s about using the LLM for generation, not final decision-making authority.
Dev: And looking at their evaluation results, they show that single-shot LLM+VAL systems already beat every direct LLM baseline and the NL-PDDL symbolic floor on Plain, confirming that the backbone is strong even before adding the evolution part.
Rosa: They also have some interesting comparisons with other models; for instance, EvoPlan with Qwen3-32B shows ninety point three percent success in their test, while a larger model like gpt-oss-120b only hits seventy-eight point four percent.
Taro: So the evolution step is definitely adding value, but it’s not a massive leap from just using a strong base model with validation, which suggests the constraint mining part is doing significant heavy lifting.
Dev: The paper also points out some limitations; they flag dependence on demonstration quality, the expressivity and tractability of the STL grammar, perception uncertainty, and sensitivity to the proposal backbone.
Rosa: Those are important caveats; we can't just assume perfect demonstrations or perfect sensor readings when we deploy this outside. We need to be careful about how expressive that signal temporal logic is for complex real-world movement.
Taro: And the need for broader hardware validation sounds like something the authors themselves recognize as a necessary next step before you trust it in a truly unstructured environment.
Dev: So, putting it all together, EvoPlan is proposing this neuro-symbolic framework where an LLM proposes plans, but those plans are checked against a formally mined mobility constraint derived from demonstrations and verified through a rigorous validation cascade.
Rosa: The authors call this EvoPlan: Evolutionary Neuro-Symbolic Robot Planning with Spatio-Temporal Guarantees because it bridges the gap between the fluency of language models and the safety guarantees needed for physical systems.
Taro: It seems like their main contribution is establishing a way to turn unstructured demonstration data into a formal, actionable safety contract that an LLM can then help navigate during planning.
Dev: The implication is that we can move towards more reliable AI planning systems where the neural part does the creative heavy lifting of proposing paths, but the symbolic and formal parts enforce the hard physical limitations.
Rosa: So what this means for us listening, it means we’re seeing a direction where these models aren't just generating text or images anymore; they are being integrated into loops that actively verify physical safety before anything moves.
Conclusion: Rosa: So, we're looking at this paper called "EvoPlan: Evolutionary Neuro-Symbolic Robot Planning with Spatio-Temporal Guarantees." It's about taking the ideas from large language models and putting them into a system that actually respects physical rules.
Dev: Yeah, it’s about using an LLM to suggest plans but making sure those plans don't break physical laws or safety protocols when you put them on a robot.
Rosa: The authors are trying to bridge that gap between the creative suggestion of the language model and the strict requirements of real-world robotics.
Dev: They achieve this by using demonstration data to mine a single global temporal logic constraint from that data, which acts like a hard rulebook for movement.
Rosa: It’s not just one simple rule, though; they turn that constraint extraction into a classification problem using counterfactual perturbations and an LLM violation generator.
Dev: That means the system learns how the robot *must* move at all times, applying that constraint to every single action any plan tries to use.
Rosa: Then they use this mined constraint as a shield during execution, checking waypoints against it before the robot actually moves anything.
Dev: The whole point is that if the LLM suggests something risky, the loop catches it immediately and forces a replan with a fresh suffix to fix that specific part.
Rosa: This framework aims to make AI planning more reliable by treating neural models as generators of candidates rather than final authority on what's safe.
Dev: It’s about building a system where the LLM proposes, but the formal verification and the temporal constraints provide the necessary safety case.
Rosa: The paper shows that even with this setup, it stays robust when it runs in open-world scenarios where the goal vocabulary doesn't perfectly match what's in the action model.
Dev: It’s interesting because they also show how graceful degradation happens when things get crowded on a robot, trading a little bit of clearance for avoiding a total system collapse.
Rosa: So, this work suggests we can start trusting these models with planning tasks more confidently if we give them that kind of formal, spatiotemporal guardrail.
Dev: And it opens up the question: how robust is this constraint mining procedure when the real-world demonstrations are messy or incomplete?
Bhavya Sai Nukapotula, Samin Moosavi Haoze Wang, Luke Duncan, Diya Shakkottai*, Varun Murali Srinivas Shakkottai
Texas A&M University
cs.RO
Submitted: 2026-07-07
Updated: 2026-11-05
Comments: Accepted at CoRL 2026. Project page: https://nbhavyasai.github.io/evoplan/
Code: https://github.com/codelion/openevolve
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: LLM-based robot planners are fluent but cannot guarantee that their plans are executable or safe because classical PDDL planners can guarantee these properties only after the problem is fully…
Key concepts
- Signal Temporal Logic (STL) Constraint ($\Phi_{mob}$)
- This is a single global rule derived from robot demonstrations that dictates how the robot must move at all times. It synthesizes missing negative examples and acts uniformly across every mobility action in any plan, serving as a crucial safety shield.
- Evolutionary PDDL Planner
- This planner uses an LLM to propose and iteratively repair plans against a fixed problem definition (PDDL). A programmatic validator scores these candidates, and the system 'evolves' by keeping the best validated prefix, ensuring monotonic improvement in plan quality.
- Constrained Execution Loop
- Before executing a proposed plan, this loop checks every waypoint against the mined STL constraint ($\Phi_{mob}$). If a violation is found, it commits the verified part of the plan and re-invokes the planner to generate a corrected suffix.
Terminology
Summary
LLM-based robot planners are fluent but cannot guarantee that their plans are executable or safe because classical PDDL planners can guarantee these properties only after the problem is fully specified and they make poor use of an LLM’s ability to read context and repair plans. p. 0 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
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. 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.Where the paper says it
Where the paper says it
System Overview
The framework involves an offline procedure that mines a single global Signal Temporal Logic (STL) constraint on mobility from demonstration data and an evolutionary PDDL planner where an LLM proposes and repairs plans against a fixed PDDL problem P and domain D. Finally, a constrained execution loop checks the plan's waypoints against the mined constraint Φmob before execution. All LLM calls use a locally-hosted open-weight model so the pipeline can be deployed on-robot with no cloud dependency. p. 0 This paper presents a
neuro-symbolic framework with three parts. All LLM calls use a locally-hosted open-weight model, so
the pipeline can be deployed on-robot with no cloud dependency.Where the paper says it
How the training signal is built
The offline constraint extraction procedure mines a single global Signal Temporal Logic (STL) constraint Φmob from demonstration data. This procedure synthesizes missing negative class with counterfactual perturbations and an LLM violation generator to turn constraint mining into a supervised classification problem. p. 1 Because every recorded trace is by construction acceptable behaviour, the
data is one-class. We synthesise the missing negative class with counterfactual perturbations and an LLM
violation generator, turning constraint mining into a supervised classification problem. We synthesise the missing negative class with counterfactual perturbations and an LLM
violation generator, turning constraint mining into a supervised classification problem. The constraint is
global: it is not attached to a particular action schema but encodes how the robot must move at all times
(bounds on speed, clearance, social spacing), so it applies uniformly to every mobility action any plan
invokes. The procedure
recovers codified rules (e. , stopping at red lights, mined from nuPlan driving logs) or population
preferences (e.Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The evolutionary PDDL planner uses an LLM to propose and repair plans against a fixed PDDL problem P and domain D. A programmatic validator cascade scores each candidate, and elitism on validator feedback grows the validated prefix of the plan monotonically over iterations. p. 1 Rather than calling a classical search procedure,
an LLM proposes and repairs plans against a fixed PDDL problem P and domain D, and a programmatic
validator cascade scores each candidate. Elitism on validator feedback grows the validated prefix of the
plan monotonically over iterations, without any explicit suffix-only mutation. When wk violates
N at runtime, the verified prefix a1 . ak−1 is committed, the resulting world state becomes the new initial
state, and the planner is re-invoked to produce a fresh suffix-replacing plan.Where the paper says it
Where the paper says it
Evaluation Results
The system was evaluated on open-world ALFWorld Text where it outperforms strong baselines and stays robust when the goal vocabulary does not match the action-model vocabulary. The mined constraint is evaluated as a shield over a Qwen2. p. 1 We evaluate the planner on open-world ALFWorld Text, where it outperforms strong baselines and stays
robust under goal/action-vocabulary misalignment (Sec. We evaluate the mined constraint as a shield
over a Qwen2. We evaluate the mined constraint as a shield
over a Qwen2. 5-VL driver on Bench2Drive and two discrete-action navigation policies on HA-VLN-CE
(Sec. 2 at clearance 0. The loop therefore degrades gracefully under crowded
operation: it trades clearance margin for an isolated failure rather than collapsing wholesale (Table 3).Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Limitations
The framework treats neural models as generators of candidate plans rather than sources of unchecked authority. Limitations include dependence on demonstration quality, the expressivity and tractability of the STL grammar, perception uncertainty, sensitivity to the proposal backbone, and the need for broader hardware validation. p. 9 37
candidates rather than sources of unchecked authority. Limitations include dependence on demonstration
quality, the expressivity and tractability of the STL grammar, perception uncertainty, sensitivity to the
proposal backbone, and the need for broader hardware validation. Limitations include dependence on demonstration
quality, the expressivity and tractability of the STL grammar, perception uncertainty, sensitivity to the
proposal backbone, and the need for broader hardware validation. Future work will extend the framework
to richer multi-robot tasks, stronger statistical guarantees for learned constraints, and per-skill constraint
families beyond global mobility.Where the paper says it
Where the paper says it
References
Bai, S. et al. (2025). Qwen2.5-VL technical report. p. 9 References
Bai, S. 5-VL technical report. 5-VL technical report. https://arxiv. 5-VL technical report. https://arxiv. 5-VL technical report. https://arxiv. https://arxiv. org/abs/2502.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Caesar, H., Kabzan, J. p. 9 org/abs/2502. , Kabzan, J. , Kabzan, J. , Wolff, E. , Wolff, E. , Fletcher, L. , Fletcher, L. , Beijbom, O. , Beijbom, O. , and Omari,
S. , and Omari,
S. nuPlan: A closed-loop ML-based planning benchmark for autonomous vehicles.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
codelion (2024). OpenEvolve: LLM-driven evolutionary program search. p. 9 codelion (2024). OpenEvolve: LLM-driven evolutionary program search. OpenEvolve: LLM-driven evolutionary program search. https://github.Where the paper says it
Where the paper says it
Donze, A. and Maler, O. p. 2 Temporal logic learning, evolutionary search, and robot safety. STL provides a language for temporal properties over real-valued signals (Maler and Nickovic, 2004), and robustness semantics support
quantitative monitoring (Donze and Maler, 2010). and Maler, O. Robust satisfaction of temporal logic over real-valued signals.Where the paper says it
Where the paper says it
Fan, J., Xu, W., Liu, M., Sokolsky, O. p. 9 In Formal
Modeling and Analysis of Timed Systems, pages 92–106. , Sokolsky, O. , Sokolsky, O. , and Kong, F. , and Kong, F. SafeGen-LLM: Enhancing safety generalization in task planning for robotic systems. SafeGen-LLM: Enhancing safety generalization in task planning for robotic systems. arXiv preprint arXiv:2602.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Hao, Y., Chen, Y., Fan, C., and Zhang, Y. p. 9 arXiv preprint arXiv:2602. , and Zhang, Y. , and Zhang, Y. Simulation to rules: A Dual-VLM framework for formal
visual planning.Where the paper says it
Where the paper says it
Helmert, M. (2006). The fast downward planning system. p. 9 Helmert, M. The fast downward planning system.Where the paper says it
Howey, R., Long, D., and Fox, M. p. 9 In Journal of Artificial Intelligence Research,
volume 26, pages 191–246. , and Fox, M. , and Fox, M. VAL: Automatic plan validation, continuous effects and mixed
initiative planning using PDDL.Where the paper says it
Where the paper says it
Jia, X. et al. (2024). Bench2Drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. p. 9 In IEEE International Conference on Tools with Artificial Intelligence. Bench2Drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. https://arxiv. org/abs/2406. https://arxiv. org/abs/2406.Where the paper says it
Where the paper says it
Where the paper says it
Karnan, H., Nair, A., Xiao, X., Warnell, G. p. 9 org/abs/2406. , Warnell, G. , Warnell, G. , Toshev, A. , Toshev, A. , Biswas, J. , Biswas, J. , and Stone, P. , and Stone, P. Socially
compliant navigation dataset (SCAND): A large-scale dataset of demonstrations for social navigation. Socially
compliant navigation dataset (SCAND): A large-scale dataset of demonstrations for social navigation. arXiv preprint arXiv:2203.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Liu, X., Toroghi, A., Liang, J., Courtis, D., Li, R., Pesaranghader, A., Kim, J., Sadhu, T., Jeon, H., and Sanner, S. (2026). Natural language PDDL (NL-PDDL) for open-world goal-oriented commonsense regression planning in embodied AI. p. 2 VLMFP generates PDDL
domain and problem files from visual observations using a dual-VLM loop, then checks symbolic execution
against simulated outcomes (Hao et al. NL-PDDL replaces rigid symbolic predicates with naturallanguage counterparts and uses lifted regression with commonsense entailment to handle open-world
goal-action misalignment (Liu et al.Where the paper says it
Maler, O. and Nickovic, D. p. 2 Temporal logic learning, evolutionary search, and robot safety. STL provides a language for temporal properties over real-valued signals (Maler and Nickovic, 2004), and robustness semantics support
quantitative monitoring (Donze and Maler, 2010). Temporal logic learning, evolutionary search, and robot safety. STL provides a language for temporal properties over real-valued signals (Maler and Nickovic, 2004), and robustness semantics support
quantitative monitoring (Donze and Maler, 2010).Where the paper says it
Where the paper says it
McDermott, D., Ghallab, M., Howe, A., Knoblock, C., Ram, A., Veloso, M., Weld, D., and Wilkins, D. (1998). PDDL–the planning domain definition language. p. 0 We illustrate the full pipeline via demonstrations using the Gazebo simulator. Keywords: Neuro-symbolic planning, planning domain description language (PDDL), signal temporal
logic (STL), safety guarantees, constrained control
1 Introduction
LLM and vision-language models can transform natural instructions, images, and demonstrations into
plausible action proposals, but a pure LLM-based plan does not by itself guarantee executability or safety.Where the paper says it
Novikov, A., Vu, N., Eisenberger, M., Dupont, E., Huang, P.-S., Wagner, A. Z., Shirobokov, S., Kozlovskii, B., Ruiz, F. J. R., Mehrabian, A., Kumar, M. P., See, A., Chaudhuri, S., Holland, G., Davies, A., Nowozin, S., Kohli, P., and Balog, M. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. p. 2 , 2025), motivating our constraint-identification stage. AlphaEvolve demonstrates that LLM-guided evolutionary search can improve machine-gradable symbolic
artifacts through automatic evaluation (Novikov et al.Where the paper says it
Ravichandran, Z., Cladera, F., Prabhu, A., Hughes, J., Murali, V. p. 0 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
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. Complementary work on grounded heterogeneous teams (Ravichandran et al. , 2025) and robot
constitutions (Sermanet et al. SafeGen-LLM: Enhancing safety generalization in task planning for robotic systems. arXiv preprint arXiv:2602. SafeGen-LLM: Enhancing safety generalization in task planning for robotic systems. arXiv preprint arXiv:2602.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Sermanet, P., Majumdar, A., Irpan, A., Kalashnikov, D., and Sindhwani, V. (2025). Generating robot constitutions and benchmarks for semantic safety. p. 2 , 2025) and robot
constitutions (Sermanet et al. , 2025) highlights the need for closed-loop grounding and semantic safety in
real robot systems.Where the paper says it
Shi, G., Wu, Y., Kumar, V., and Sukhatme, G. S. (2025). PIP-LLM: Integrating PDDL-integer programming with LLMs for coordinating multi-robot teams using natural language. p. 2 NL-PDDL replaces rigid symbolic predicates with naturallanguage counterparts and uses lifted regression with commonsense entailment to handle open-world
goal-action misalignment (Liu et al. PIP-LLM uses PDDL for team-level task decomposition and
integer programming for robot-level allocation in multi-robot coordination (Shi et al.Where the paper says it
Soroka, E., Sinha, R., and Lall, S. (2025). Learning temporal logic predicates from data with statistical guarantees. p. 2 STL provides a language for temporal properties over real-valued signals (Maler and Nickovic, 2004), and robustness semantics support
quantitative monitoring (Donze and Maler, 2010). Recent methods learn temporal-logic predicates from
data with statistical guarantees (Soroka et al.Where the paper says it
Wijmans, E., Kadian, A., Morcos, A., Lee, S., Essa, I., Parikh, D., Savva, M., and Batra, D. (2020). DD-PPO: Learning near-perfect PointGoal navigators from 2. p. 6 The constraint shields two discrete-action policies on HA-VLN-CE (Wu et al. ,
2024): a pretrained PointNav (DD-PPO (Wijmans et al.Where the paper says it
Wu, F. et al. (2024). Human-aware vision-and-language navigation: Bridging simulation to reality with dynamic human interactions. p. 5 ak−1 is committed, the
state is refreshed from the resulting world, and the planner is re-invoked from the updated initial state. The bridge is implemented as a ROS 2 adapter that uses a differential-drive robot model (a simulated Jackal
in our Gazebo demonstration), local distance fields, human tracking, and terrain labels; the navigation
layer is realised by a trajectory planner whose global and local stages produce a waypoint sequence.Where the paper says it
Table 8
Method Total tokens (N=134) SR
Fast Downward (upper bound) - 100%
GPT-5.5 (LLM+VAL loop) 2,770,945 100%
EvoPlan (Qwen3-32B) 10,349,634 90.3%
EvoPlan (gpt-oss-120b) 17,071,316 78.4%
Table 9
Single-shot LLM+VAL (k=1, no evolution) 184,178 86.7
EvoPlan (feedback-conditioned evolution, iters=10) 273,400 99.7
Single-shot Qwen3-32B already beats every direct-LLM baseline and the NL-PDDL symbolic floor on Plain, confirming the backbone is strong
Table 7
Plain Mystery Randomized Misalign.
GPT-4o 982,602 34%
Gemini-2.0 Flash 928,024 18%
LLaMA-3.1 963,565 44%
NL-PDDL (Liu et al., 2026) - 70%
Fast Downward (upper bound) - 100%
EvoPlan (Qwen3-32B) 99.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Constraint mining from demonstrations can recover codified rules like
stop at red lights
or population preferences likesocial-navigation comfort,
allowing for a single global Signal Temporal Logic constraint Φmob that shields a vision-language driving policy on Bench2Drive or two discrete-action navigation policies on HA-VLN-CE. -
An evolutionary PDDL planner can be implemented where an LLM proposes and repairs plans against a fixed PDDL problem, and a programmatic validator cascade scores each candidate, ensuring the validated portion of the plan grows monotonically over iterations.
-
A constrained execution loop can be integrated so that every action is mapped into a waypoint sequence that is checked against Φmob before execution; if it violates Φmob, the verified prefix is committed and the planner re-invokes on an updated state, enabling robust performance in dynamic environments like Gazebo.
-
The system can handle goal/action-vocabulary misalignment by using a one-time LLM call to derive a predicate-symbol map µ that is cached and reused during the evolutionary planning process, allowing it to compose cleanly with open-world belief growth.
-
The constraint mining procedure can be adapted to recover preference rules from teleoperation data, which can then be used as a shield for discrete-action policies where the learned constraint compresses heavy unsafe tails in per-episode safety distributions by removing 82% of avoidable contacts on PointNav and 93% on the MLP.
-
The constrained execution loop can generate failure facts like route-blocked or no-safe-approach, which are then consumed by the planner to choose alternate actions rather than repeatedly attempting a failed one, leading to graceful degradation under crowd density where the loop trades clearance margin for isolated failures rather than collapsing wholesale.
Abstract
LLM-based robot planners are fluent but cannot guarantee that their plans are executable or safe. Classical PDDL planners can guarantee symbolic validity, but only once the problem is fully specified, and they make poor use of an LLM's ability to read context and repair plans. We present EvoPlan, a neuro-symbolic framework whose guarantees are symbolic validity and satisfaction of a constraint learned from data. The planner works with locally-hosted open-weight models, so it does not require a cloud service. EvoPlan has three parts. First, an offline procedure that mines a single global Signal Temporal Logic (STL) mobility constraint from demonstrations. Perturbed demonstrations supply the missing negatives without encoding any rule, and evolutionary search fits the constraint, which captures driving rules or social-navigation preferences within a modelled taxonomy of norms. We use the mined constraint to shield driving and navigation policies on Bench2Drive and 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. On open-world ALFWorld and NL-PDDL Blocksworld, it approaches classical planning when the goal vocabulary agrees with the action model, and it stays robust when the two disagree and classical planning fails. Third, a constrained execution loop: Nav2 uses the costmap to ground the plan into waypoints, which are checked against the mined constraint. Violations and execution failures update the PDDL problem for re-planning. We demonstrate the full pipeline in Gazebo and on a Clearpath Jackal robot. More details and videos are on the project page: https://nbhavyasai.github.io/evoplan/
Sources
- Qwen2.5-VL Technical Report
- SafeGen-LLM: Enhancing Safety Generalization in Task Planning for Robotic Systems
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving
- Socially Compliant Navigation Dataset (SCAND): A Large-Scale Dataset of Demonstrations for Social Navigation
- Heterogeneous Robot Collaboration in Unstructured Environments with Grounded Generative Intelligence
- PIP-LLM: Integrating PDDL-Integer Programming with LLMs for Coordinating Multi-Robot Teams Using Natural Language
- Human-Aware Vision-and-Language Navigation: Bridging Simulation to Reality with Dynamic Human Interactions
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation