Projects

Ascendr

Teaching an agent to climb

Role
Personal project
When
Mar – Apr 2025
Built with
Python, Gymnasium, Matplotlib, reinforcement learning
  • 4discrete actions, grab, release, shift left, shift right

Ascendr is a climbing environment for reinforcement learning, built on Gymnasium. A climber moves up a wall by interacting with holds, and the environment is designed to plug straight into algorithms like DQN. It can also be played by hand.

The agent chooses from four discrete actions: grab a hold, release a hold, or shift an arm left or right. It observes its torso and arm positions, the holds available on the wall, the target hold, and its distance to that target.

The environment renders the climber’s state in real time with Matplotlib, so training runs can be watched as they happen.

NextRL Agent for Tron