Source-linked AI summary
SLoSa: Small, Low-Power, and Safe Foundation Models for Robotic Task Planning
Weizhe Xu, Jialiang Fan, Essien Taylor, Zijun Wang, Russell Joseph, Fanxin Kong
TL;DR
Safety-constrained robotic planning must produce executable, goal-reaching plans while satisfying trajectory-level requirements, but classical planners use rigid symbolic inputs and general-purpose LLMs can be costly to deploy locally. SLoSa transfers planner and verifier feedback into compact LLM planners through safety-aware data construction, supervised fine-tuning, and verifier-guided reinforcement learning. The authors report consistent improvements across model sizes, domains, and input representations, alongside evaluations in household simulations, on edge platforms, and with a desktop robotic arm.
Problem
Robotic task planners must reach goals while satisfying trajectory-level safety constraints, yet classical planners require rigid symbolic inputs and general-purpose LLMs can be too large for resource-limited robots.
Source: §I · p. 1Method
SLoSa constructs safety-aware training data with planners and formal verifiers, then applies supervised fine-tuning and verifier-guided reinforcement learning to compact LLM planners.
Source: §I · pp. 1–2Results
Safety-aware post-training consistently improves results across model sizes, task domains, and input representations, with evaluations extending to household benchmarks, local inference, and a physical robotic arm.
Source: §I · p. 2Takeaways & Limitations
SLoSa provides a practical way to combine flexible LLM-based plan generation with symbolic planning and formal verification across PDDL and additional robotic task settings.
Source: §VII · p. 8The evaluation covers high-level constraints derived from symbolic action models, not perception errors, motion-level failures, or unmodeled physical risks; the desktop-arm test uses one workcell and fixed low-level skills.
Source: §VI · p. 8
Abstract
from arXiv · showhide
Safe robotic task planning requires planners that can reliably achieve task goals while satisfying safety constraints. Classical planners search over explicit symbolic task representations and can provide formal guarantees with respect to encoded safety constraints, but require rigid input formats. Their computational cost also increases rapidly with the search space. In contrast, large language models (LLMs) offer flexible task interfaces and strong generalization capabilities, but they are large, power-hungry, and pre-trained with little robot-specific safety data and knowledge. Thus, they can be difficult to deploy locally on robotic systems with limited resources. To address these limitations, we present SLoSa, a neuro-symbolic compression and fine-tuning framework for training foundation-model-based robotic task planners with Small size, Low power demand, and Safety awareness. The framework first constructs a safety-aware dataset using classical planners, an advanced teacher LLM, and formal verifiers, followed by two training stages: supervised fine-tuning and reinforcement learning (RL) fine-tuning guided by a verification-based reward machine to enforce safety alignment. We evaluate SLoSa across four Planning Domain Definition Language (PDDL) domains, two household simulation environments, local edge platforms, and a desktop robotic arm. Across the PDDL tasks, verifier-guided post-training raises the success rate of compact 4B-14B planners to 70.5-76.0%, while INT4 deployment on a Jetson Orin requires 17.85-26.41 W of average power. Results in household simulators and on the physical robot further demonstrate the applicability of SLoSa across diverse robotic task settings.
I. INTRODUCTION
SLoSa addresses the need for robotic planners that combine flexible task inputs with compact deployment and plans that satisfy explicit trajectory-level safety constraints. It transfers planner and verifier feedback into compact LLMs through safety-aware data construction and post-training, with evaluations spanning PDDL, household simulations, edge platforms, and a robotic arm.
I. INTRODUCTION
- Safety-constrained planning requires executable, goal-reaching plans whose intermediate states satisfy explicit trajectory-level requirements.Requirements can include ordering, mutual exclusion, and resource constraints.Source: §I · p. 1
- Classical planners offer formal safety criteria but can face rapidly growing search spaces and require rigid symbolic inputs and engineered domain specifications.Learned policies can amortize planning, but outputs lack formal guarantees without an explicit specification and verifier.Source: §I · p. 1
- LLMs offer flexible task interpretation, but general-purpose models can be too large for resource-limited robots, and verifier-only repair leaves the base planner unchanged.Repeated failed generations increase token use, latency, and computation.Source: §I · p. 1
- SLoSa uses behavior-changing safety-aware data, compression-aware supervised fine-tuning, and verifier-guided GRPO to improve compact LLM planners.Formal feedback changes the planner during post-training, while inference combines local plan generation with verifier feedback.Source: §I · pp. 1–2
- The evaluation spans four PDDL domains and two adapted household benchmarks, with local inference tests and a vision-grounded desktop 6-DoF arm demonstration.The PDDL domains cover manipulation, transportation, object transfer, and tool use.Source: §I · pp. 1–2
- Safety-aware post-training consistently improves results across model sizes, task domains, and input representations, with evaluations also covering local edge inference and physical robotic-arm execution.The evaluation includes four PDDL domains and separately trained planners for VirtualHome and ALFWorld.Source: §I · pp. 1–2
II. RELATED WORK
Prior approaches use symbolic planning, learned policies, or LLMs with formal grounding and runtime verification, each with different trade-offs in specification, scaling, and repair. SLoSa uses symbolic tools as supervision and defines planning and verification outcomes against explicit goals and safety constraints.
II. RELATED WORK
- Classical planners provide explicit, interpretable task semantics but require labor-intensive specifications and can face exponentially growing search spaces; learned policies are generally tied to training distributions.SLoSa uses symbolic planners and validators primarily as training supervision, with a compact neural model generating deployment plans.Source: §II · p. 2
- SLoSa differs from verifier-loop approaches by using formal validation during post-training to improve compact planners’ first-pass plans, while retaining the option of verifier-based repair at deployment.Validator-accepted plans retain formal guarantees, and the repair loop can reduce regeneration tokens.Source: §II · p. 2
- Compression reduces LLM memory and adaptation costs, but does not itself optimize task-level safety; SLoSa studies transferring verifier knowledge into compact planners.This complements work on safety generalization using verified PDDL supervision.Source: §II · p. 2
III. PRELIMINARIES
- The planning problem is represented as P = (S, A, s0, G, C), comprising symbolic states, actions, an initial state, a goal, and safety constraints.Actions have preconditions and effects and represent high-level skills such as pick, move, or place.Source: §III · p. 2
- The verifier assigns each output exactly one category according to the first failed check, distinguishing format, precondition, safety, and goal failures from success.Safety is checked only after a complete executable trajectory is obtained.Source: §III · p. 2
- SLoSa defines safety awareness as compliance with the formal task specification, while low-power operation remains an empirical deployment objective.Its goal is a compact LLM planner producing safety-compliant plans with low deployment and inference cost.Source: §III · p. 2
IV. SLOSA
SLoSa combines verified training examples, compression-aware supervised fine-tuning, and verifier-guided reinforcement learning to train compact quantized planners. Its data construction selects constraints that change the required behavior relative to a baseline plan while preserving a verified alternative.
IV. SLOSA
- SLoSa trains compact quantized LLM planners using verified plans for supervised fine-tuning and verifier outcomes plus planning-progress signals for verifier-guided GRPO.Inputs may be PDDL, structured text, or natural language, and outputs are grounded action sequences.Source: §IV · pp. 2–3
A. Safety-Aware Training Data Construction
- The data-construction stage uses four robotics-inspired PDDL domains covering manipulation, transportation, object transfer, and tool use.The domains are Blocksworld, Ferry, Grippers, and Spanner.Source: A. Safety-Aware Training Data… · p. 3
- A retained instance requires a successful baseline plan, a successful constrained alternative, and a baseline plan that violates the added constraint.This makes the constraint behavior-changing relative to the selected baseline, not proof that the formal specification captures every real-world hazard.Source: A. Safety-Aware Training Data… · p. 3
- Difficulty levels are based solely on verified plan length, so they represent effective planning horizon rather than problem-size parameters.The dataset also renders constrained problems as PDDL and structured or narrative natural-language prompts.Source: A. Safety-Aware Training Data… · p. 3
- Verified constrained plans become positive SFT examples, while 20% of constrained problems receive additional teacher-generated plans retained only when formally verified as successful.GRPO uses the constrained problems without target plans.Source: A. Safety-Aware Training Data… · p. 3
B. Compression-Aware Supervised Fine-Tuning
SFT gives compact planners a strong starting point for producing parseable, executable plans close to verified constrained plans. It uses verifier-accepted plans as targets while adapting quantized models with parameter-efficient adapters.
B. Compression-Aware Supervised Fine-Tuning
- SFT initializes compact planners to generate parseable, executable plans close to verified constrained plans before GRPO.The pretrained models lack the required planning format, action semantics, and safety-constrained behavior.Source: B. Compression-Aware Supervised… · p. 3
- SFT examples pair prompts from constrained planning problems with verifier-accepted plans, and only verified plans serve as positive targets.The target output is a grounded action sequence with one action per line.Source: B. Compression-Aware Supervised… · p. 3
- Compact LLMs are fine-tuned with quantization and parameter-efficient adapters, with only adapter parameters trained.The base model is loaded in low precision.Source: B. Compression-Aware Supervised… · p. 3
C. Verifier-Guided GRPO
GRPO refines the SFT planner using verifier outcomes as rewards, with classical-planner distance distinguishing progress within a failure category. A difficulty curriculum and relative comparisons among sampled plans guide updates.
C. Verifier-Guided GRPO
- GRPO uses a formal verifier as a programmatic reward function, avoiding a separate reward model or critic.For each prompt, the policy samples candidate plans and the verifier assigns each one a mutually exclusive outcome and scalar reward.Source: C. Verifier-Guided GRPO · p. 3
- Format errors receive −1 and successful plans receive 1, while precondition violations, safety violations, and goal failures receive increasingly less-negative reward intervals.This ordering treats safety violations as more severe than executable plans that fail only to reach the goal.Source: C. Verifier-Guided GRPO · p. 4
- For failures, classical-planner distance measures progress toward the goal and orders candidates within the verifier-defined reward interval.If the residual planning problem cannot be solved within the time limit, the lowest progress value is assigned.Source: C. Verifier-Guided GRPO · p. 4
- GRPO compares normalized advantages within prompt groups, excludes groups with identical rewards, and gradually shifts training from shorter to longer verified plans.When a nominal step has no active group, optimizer and learning-rate updates are skipped; the curriculum begins with shorter plans before increasing medium- and hard-plan proportions.Source: C. Verifier-Guided GRPO · p. 4
V. EXPERIMENTS
The evaluation spans PDDL domains, household simulations, a robotic arm, local deployment costs, and pipeline ablations.
V. EXPERIMENTS
- SLoSa is evaluated in PDDL domains, household simulation benchmarks, and on a real-world robotic arm.The evaluation also measures local deployment cost and ablates the safety-aware post-training pipeline.Source: §V · p. 4
A. Experimental Setup
The experiments use disjoint planning-problem splits across four PDDL domains and separately trained household-benchmark planners, then assess post-training performance across models, domains, and input representations. On the 200-instance PDDL evaluation, GRPO-refined compact models reach 70.5–76.0% success, while Qwen3-4B results are comparable across three input formats.
A. Experimental Setup
- Each of four PDDL domains contributes 250 constrained problems, represented as PDDL and either structured or narrative natural-language prompts, with splits disjoint by underlying problem.The PT, SFT, and GRPO models share the same test split and output format; evaluation uses greedy decoding without intermediate chain-of-thought output.Source: A. Experimental Setup · pp. 4–5
- Household evaluations use separately trained Qwen3-4B planners on fully observed, non-interactive adaptations of VirtualHome and ALFWorld.ALFWorld-FO-Safe exposes the complete symbolic state and is not the standard partially observed, interactive ALFWorld protocol.Source: A. Experimental Setup · p. 5
B. Planning Performance
- 70.5%, 74.0%, and 76.0% success rates are achieved by GRPO-refined 4B, 8B, and 14B models, respectively, on the 200-instance PDDL evaluation.SFT raises success from 0% to 59.0–67%; directly prompted GPT-5.6 Sol and Claude Sonnet 5 achieve 26.5% and 31.0%, respectively.Source: B. Planning Performance · p. 5
- Qwen3-4B post-training gains are consistent across four PDDL domains, and its GRPO-refined model performs comparably with PDDL, structured natural language, and narrative natural language inputs.Blocksworld remains the most challenging domain.Source: B. Planning Performance · p. 5, Figure 2 · p. 5
C. Contribution of Safety-Aware Post-Training
The ablation compares equally sized positive-only SFT datasets to isolate how safety-aware data composition affects planning. Constraint-aware supervision improves success over unconstrained SFT, and behavior-changing examples further reduce unsafe goal-reaching plans.
C. Contribution of Safety-Aware Post-Training
- The ablation isolates data composition by comparing unconstrained plans, a balanced mixture, and behavior-changing instances using SFT alone.Behavior-changing instances invalidate the nominal plan while preserving a verified alternative.Source: C. Contribution of Safety-Aware… · p. 5
- Success requires an executable, goal-reaching plan that satisfies constraints, while goal reach ignores constraint compliance and conditional safety violation measures violations among goal-reaching outputs.These metrics distinguish completing a task from complying with its safety constraints.Source: C. Contribution of Safety-Aware… · p. 5
- Constraint-aware supervision substantially improves success over unconstrained positive-plan SFT.The variants use equally sized positive-only datasets and no contrastive negative examples.Source: C. Contribution of Safety-Aware… · p. 5
- Behavior-changing training examples further reduce unsafe goal-reaching plans, indicating that filtering teaches responses to constraints rather than only goal-reaching solutions.The comparison uses SFT alone, without contrastive negative examples or GRPO.Source: C. Contribution of Safety-Aware… · pp. 5–6
D. Applicability Across Planning Benchmarks
SLoSa’s post-training gains extend to separately trained household-simulation planners, where SFT improves success and verifier-guided GRPO adds further gains. The section also reports local edge deployment, runtime scaling, and reduced verification-guided repair costs.
D. Applicability Across Planning Benchmarks
- Across both household benchmarks, SFT improves planning success over the pretrained model, and verifier-guided GRPO adds a further gain.The trends hold across different task representations and execution environments.Source: D. Applicability Across Planning… · p. 6
- Figure 3 contrasts baseline plans that violate added constraints with safety-satisfying plans for the same simulator tasks.The examples illustrate constraint-driven changes in planning behavior in VirtualHome-Safe and ALFWorld-FO-Safe.Source: D. Applicability Across Planning… · p. 6
E. Edge Deployment and Repair Cost
- On Jetson Orin, compact INT4 planners use 3.89–11.50 GB peak memory, run at 3.55–5.01 tokens/s, and draw 17.85–26.41 W on average.All three compact planners fit and run locally on both the RTX 4080 and Jetson Orin under the reported setup.Source: E. Edge Deployment and Repair Cost · p. 6
- Fast Downward and OPTIC show increasingly variable runtimes as Blocksworld instances grow, with many larger instances reaching the 300-s timeout, while GPT-5.6 Sol remains comparatively stable.The comparison covers 41 instances with 10–50 blocks; some larger instances solve faster because search difficulty depends on the initial and goal configurations.Source: E. Edge Deployment and Repair Cost · p. 6
- Stronger SLoSa post-training reduces SafePilot repair iterations and token usage while improving final success under a three-round repair limit.Only plans passing all verifier checks are accepted.Source: E. Edge Deployment and Repair Cost · p. 6
F. Desktop Robot-Arm Demonstration
The desktop-arm demonstration combines image- and language-based task specification, symbolic planning, and verifier-gated execution. Across 30 Blocksworld trials, it demonstrates practical feasibility while distinguishing verifier rejection from physical execution failures.
F. Desktop Robot-Arm Demonstration
- In the simulator examples, baseline plans complete the tasks but violate added ordering constraints, while lower-row plans satisfy them.The examples concern symbolic safety compliance, not physical collision safety.Source: Figure 3 · p. 7
- A vision-language model converts initial- and goal-state images plus natural-language safety constraints into a safety-constrained PDDL3 problem for Qwen3-4B planning.The model extracts objects, initial state, goal predicates, and safety constraints.Source: F. Desktop Robot-Arm Demonstration · pp. 6–7
- A formal verifier checks each candidate plan before execution, returns rejection feedback for revision, and passes only accepted plans to the controller.Symbolic actions are mapped to predefined manipulation skills, reducing the risk of invalid high-level sequences reaching the robot.Source: F. Desktop Robot-Arm Demonstration · p. 7
- Across 30 Blocksworld trials, the robot-arm experiment demonstrates the feasibility of a complete high-level planning and verification pipeline in a real-world environment.Verifier acceptance and physical completion are reported separately, and rejected plans are not executed.Source: F. Desktop Robot-Arm Demonstration · p. 7
- Low-level physical failures can occur after verifier acceptance when manipulation precision causes stacking errors or stacks exceed the arm’s reachable workspace.The reported physical failure in the representative execution is attributed to limited manipulation precision, not a symbolic-constraint violation.Source: F. Desktop Robot-Arm Demonstration · p. 7, Figure 5 · p. 8
VI. DISCUSSION AND LIMITATIONS
SLoSa transfers formal planning feedback into compact LLMs while preserving flexible task interfaces. The authors limit their safety claims to high-level symbolic constraints and report that perception errors, motion-level failures, and unmodeled physical risks remain outside the evaluation.
VI. DISCUSSION AND LIMITATIONS
- SLoSa uses symbolic planning and formal verification for training feedback rather than restricting deployment inputs to symbolic formats.The resulting planners accept formal language, structured natural language, and narrative natural language.Source: §VI · p. 8
- The evaluation covers high-level constraints derived from symbolic action models and verifier outcomes, not perception errors, motion-level failures, or unmodeled physical risks.The desktop-arm evaluation uses one workcell and fixed low-level skills; long-horizon domains such as Blocksworld remain challenging.Source: §VI · p. 8
VII. CONCLUSION
- The conclusion presents SLoSa as a way to combine flexible LLM-based plan generation with symbolic planning and formal verification.It reports improved executable, goal-reaching, and safety-compliant planning across model sizes, domains, and input formats.Source: §VII · p. 8