LIFTER: Deep RL for Fall-Recovery Control of Quadruped Robots
deep-rl
reinforcement-learning
robotics
quadruped
isaac-gym
Deep reinforcement learning for quadruped robot fall recovery using Isaac Gym + PPO
1 Overview
Deep reinforcement learning research project for fall-recovery control of quadruped robots on non-flat terrain. LIFTER (Legged Incline Fall-recovery Technique for Enhanced Recovery).
2 Tech Stack
| Category | Technology |
|---|---|
| Simulation | Isaac Gym (NVIDIA), PyBullet |
| RL Algorithm | PPO (Proximal Policy Optimization) |
| Neural Network | CENet (Context Encoder Network) |
| Framework | PyTorch |
| Hardware | NVIDIA GPU (CUDA) |
| Language | Python 3.8 |
3 Key Features
- Fallen Pose Sampling: Efficient sampling of 3,000 fallen poses using K-ACC clustering (Stage 1)
- Multi-terrain Training: Isaac Gym-based training on various terrains including flat, rough, and stairs (Stage 2)
- Context-based Recovery Learning: Fall recovery policy learning with context awareness using CENet
- Multi-robot Support: Compatible with various quadruped platforms including AiDIN-6 and AiDIN-8