A comparative study of classical and learned control strategies for cart-pole stabilisation, implemented and benchmarked in NVIDIA Isaac Sim with Isaac Lab.
Three controllers are implemented: PID (three architectures), LQR, and PPO (via curriculum learning). The project covers the full pipeline — asset construction in USD, physics parameterisation, controller design, and a structured evaluation framework.
🚧 Work in progress — curriculum training and robustness evaluation are ongoing. Results and extensions will be added as they are completed.
| GPU | NVIDIA RTX 3090 |
| CPU | AMD Ryzen 9 |
| OS | Ubuntu 22.04 |
| Simulator | NVIDIA Isaac Sim 5.1 + Isaac Lab |
| RL Framework | skrl (PPO, 4096 parallel envs) |
| Language | Python |
This project is part of a broader self-directed programme in robotics and control, with two main motivations:
- Skill building — hands-on experience with GPU-accelerated simulation, classical control theory (PID, LQR), and reinforcement learning in a physically realistic environment.
- Community sharing — documenting the full development process openly, including design decisions, pitfalls, and results, so others getting started with Isaac Sim or control RL have a reference to learn from.
- Cart-pole asset built from scratch in Isaac Sim (USD)
- Physics validated against analytical linearised dynamics (<1% error)
- PID (single-loop, dual-loop, cascade) implemented and compared
- LQR designed and benchmarked
- PPO trained — curriculum Stages 1 & 2 (up to ±30° initialisation)
- Structured evaluation framework with CSV logging
- PPO curriculum Stages 3–5 (up to ±150°, full swing-up)
- Robustness evaluation (impulse disturbances, sensor noise, Kalman filter)
- Physical parameter sweeps (damping, mass ratio)
- SAC vs PPO comparison
- Cross-simulator benchmarking (MuJoCo, Gazebo)
- Double-pendulum extension
A technical report covering the methodology and completed results is available in Report.
K Lab — May 2026