Research map / Label-efficient and robust learning / Recognition and long-tail
Safe Planner: Empowering Safety Awareness in Large Pre-Trained Models for Robot Task Planning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Safe Planner, which adds a learned collision-risk prediction module (frozen pre-trained ViT plus LSTM, trained on simulator data) to guide a GPT-4V planner over PDDL skills, reducing collisions in simulated and real robot tasks.
Problem
Large pre-trained VLM/LLM planners for robot task planning lack real-scene grounding and can produce unsafe plans that cause collisions and damage to robots and environments.
Approach
Train a safety prediction module in simulation that estimates per-skill, per-object collision risk; convert the resulting safety matrix into natural-language rankings that guide a GPT-4V PDDL planner to clear obstacles before manipulating targets, and transfer the module zero-shot to a real Franka arm.
Key results
- In simulation (Chair Hard), Safe Planner reduces average collisions to 6.21 versus 10.50 for SayCan and 10.51 for PROGPROMPT.
- On real Franka robot, Safe Planner yields 0.2/1.0/1.0 collisions across easy/medium/hard versus 2.0/4.2/2.1 without the safety module.
- Safe Planner has lower success rates than the ablation without the safety module in several simulated settings (e.g., Table Easy 0.70 vs 0.67 is higher, but Table Hard 0.73 vs 0.80 is lower), and reports 0.7 success on real hard tasks versus 0 without SM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the safety module (ablation) | Safe Planner · simulated Table, Counter, and Chair tasks (Easy and Hard) | average collisions: 3.80 -> 2.44; 5.42 -> 4.33; 2.63 -> 2.07; 3.20 -> 1.73; 2.30 -> 1.26; 10.59 -> 6.21 | Helped |
| add the safety module (ablation) | Safe Planner · simulated Table, Counter, and Chair tasks (Easy and Hard) | average task success rate: 0.67 -> 0.70; 0.80 -> 0.73; 0.40 -> 0.40; 0.40 -> 0.40; 0.95 -> 0.90; 0.83 -> 0.82 | Mixed |
| add the safety module (ablation) | Safe Planner · real robot Easy, Medium, and Hard tasks | average collisions: 2.0 -> 0.2; 4.2 -> 1.0; 2.1 -> 1.0 | Helped |
| add the safety module (ablation) | Safe Planner · real robot Hard | average success rate: 0 -> 0.7 | Helped |
| add the safety module (ablation) | Safe Planner · real robot Easy and Medium | average success rate: 1.0 -> 1.0; 1.0 -> 1.0 | No clear effect |
| use a pre-trained ResNet50 instead of ResNet50 from scratch (ablation) | Safety prediction module · Chair Hard | collisions: 10.59 -> 6.48 | Helped |
| use a pre-trained ViT instead of ViT from scratch (ablation) | Safety prediction module · Chair Hard | collisions: 10.52 -> 6.21 | Helped |
| use pre-trained ViT instead of pre-trained ResNet50 (ablation) | Safety prediction module · Chair Hard | collisions: 6.48 -> 6.21 | Helped |
Surprising findings
- Without the safety module, the framework's collision counts are nearly the same as the baselines, and improving safety came at a small cost in simulated success rates because safer plans are longer.
Limitations
- Safety is defined only as collision count, not broader safety notions such as force or velocity limits.
- Relies on hand-crafted scripts for real-robot skills and on GPT-4V, with planner reliance on state predicates.
- Real-robot evaluation uses only 10 runs per setting and a simple pick-and-place task.
Experimental setup
Datasets: simulated Habitat 2.0 trajectories with randomized scenes ImageNet
Benchmarks: Habitat 2.0 housework tasks (Table, Counter, Chair scenes) real Franka Panda pick-and-place tasks
Metrics: average collision number task success rate
Baselines: SayCan PROGPROMPT Safe Planner w/o SM
Builds on: SayCan ProgPrompt Grounded Decoding PDDL
Closest papers by meaning
- SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning (NeurIPS 2025 · similarity 0.88)
- ABNet: Adaptive explicit-Barrier Net for Safe and Scalable Robot Learning (ICML 2025 · similarity 0.86)
- AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions (CVPR 2026 · similarity 0.85)
- Introspective Planning: Aligning Robots' Uncertainty with Inherent Task Ambiguity (NeurIPS 2024 · similarity 0.85)
- IS-Bench: Evaluating Interactive Safety of VLM-Driven Embodied Agents in Daily Household Tasks (AAAI 2026 · similarity 0.85)
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.