Skip to content
All research
Research 04Published · simulation

Hybrid RL and LLM Framework for Robotic Manipulation

Language models plan the task; reinforcement-learning policies carry out the motion.

IEEE EmergIN 2025 · published conference paper

Published paper (DOI)Open preprint

System overview

  1. 01Task in natural language
  2. 02LLM planner
  3. 03Sub-goals
  4. 04RL policy (PPO/SAC)
  5. 05Panda in PyBullet
Simplified flow of RL + LLM hybrid, following the published description.

The problem

RL policies are good at precise low-level control but struggle to generalise across task descriptions; LLMs reason well about tasks but cannot control joints. The framework splits the work between them.

Approach

  1. 01An LLM interprets the natural-language task and decomposes it into sub-goals.
  2. 02RL policies (PPO and SAC) execute each sub-goal as low-level motion.
  3. 03Evaluated in PyBullet with a simulated Franka Emika Panda on pick-and-place, object sorting and dynamic obstacle avoidance.

Evidence

Compared with an RL-only baseline in simulation, the study reports shorter task completion times and higher success and adaptability. Metric definitions and trial counts are not detailed, so these numbers are not reproduced here.

Scope and limitations

Simulation-based evaluation. Real-world robot validation and sim-to-real transfer remain future work.

Publication

2025

Conference

Published

Hybrid Framework for Robotic Manipulation: Integrating Reinforcement Learning and Large Language Models

IEEE International Conference on Emerging Technologies and Innovation for Sustainability (EmergIN 2025) · IEEE EmergIN 2025 · pp. 886–890

LLM-based task planning on top of RL-based low-level control (PPO/SAC), evaluated in PyBullet with a simulated Franka Emika Panda.

DOI for Hybrid Framework for Robotic Manipulation: Integrating Reinforcement Learning and Large Language Models (opens in a new tab)arXiv for Hybrid Framework for Robotic Manipulation: Integrating Reinforcement Learning and Large Language Models (opens in a new tab)Case study