Safety-Aware LLM Planning for Human-Robot Collaboration in Manufacturing

An end-to-end pipeline integrating LLM action planning with a Neo4j knowledge graph of ISO safety standards and VLM scene understanding, enabling compliant cobot behavior in manufacturing settings.

Manufacturing environments increasingly deploy collaborative robots (cobots) that must share workspace with human workers. Ensuring safe operation requires real-time awareness of both the physical scene and applicable safety regulations — a challenge traditional rule-based planners handle poorly. This project develops a Safety-Aware LLM (SA-LLM) pipeline that integrates large language model reasoning with a structured knowledge graph of industrial safety standards to generate compliant, context-aware robot action sequences.

System Architecture

The SA-LLM pipeline has five tightly integrated components:

1. Knowledge Graph (Neo4j). Industrial safety standards — IEC 60204-1, ISO 10218, ISO 13850, ISO/TS 15066, and ISO 12100 — are ingested into a Neo4j graph database. Documents are chunked (~500 tokens, 50-token overlap) and each chunk is embedded using sentence-transformers (all-MiniLM-L6-v2, 384-dim) for vector similarity search. The graph encodes nodes for SafetyStandard, Requirement, Constraint, and CollaborativeMode, connected by typed relationships DEFINES, REQUIRES, LIMITS, and APPLIES_TO.

2. Visual Scene Understanding (Llama 3.2-Vision, 11B). RGB images from the simulation are passed to a vision-language model that produces structured scene descriptions: object identities, spatial relationships, and human pose and distance relative to the robot — all needed to select appropriate safety rules.

3. Context-Aware Safety Retrieval (GraphRAG). A query combining task type, human distance, and scene context is embedded and used to retrieve applicable safety rules via vector similarity. This maps natural-language context (e.g., “worker at 0.3 m”) to concrete regulatory thresholds such as ISO/TS 15066 speed and force limits, without relying on brittle keyword matching.

4. LLM Action Planning (Qwen3-8B). The reasoning LLM receives the scene description, retrieved safety rules, the human command, and the available action primitive set. It outputs a structured plan specifying IS_SAFE, SPEED_LIMIT, FORCE_LIMIT, SAFETY_REASONING, and an ordered action sequence.

5. Action Primitives and Execution. Nine primitives are defined — move_to(obj), pick(obj), place(x,y,z), hand_over(), retract(), wait(s), speak(msg), stop(), adjust_speed(factor) — which the LLM composes into executable sequences. Actions are dispatched to a simulated Franka Panda (7-DOF, 855 mm reach) in PyBullet and deployed on physical hardware via ROS and the franka_ros / Franka Control Interface (FCI) stack, providing a clean sim-to-real deployment path with the same 9 primitives in both environments.

Evaluation

Four test scenarios probe different safety-critical aspects of human-robot collaboration:

Scenario Human Distance Task Safety Objective
S1 0.29 m “Give me the wrench” Enforce ISO/TS 15066 speed and force limits for close proximity
S2 0.35 m, blocking path “Pick up part” Refuse or redirect around restricted zone
S3 Transient hand contact Object handover Retrieve hand contact force limits
S4 1.01 m “Mate the parts” Assembly with appropriately relaxed constraints

The pipeline correctly retrieves and applies ISO safety constraints when grounded in the knowledge graph, producing action sequences that adapt speed and force limits based on real-time human proximity — without hardcoded rules for each scenario.

Transferable Engineering

This work directly applies to: cobot deployment in manufacturing and logistics, safety certification workflows (CE marking, ISO compliance), LLM-powered robot task planning, embodied AI systems that must reason over structured regulatory domain knowledge at runtime, and sim-to-real robot deployment pipelines that require compliant behavior in human-occupied workspaces.