a bit more about how it works:
- PPO in MuJoCo, no real-world data
- 0-4 frames of random latency, soft servos, noisy ball state
- on the arm: wrist cam -> colour blob -> MLP position -> finite-diff velocity -> policy -> joint deltas
code: https://t.co/30o8m26wOe
@dominiconorton thanks! the biggest challenge was getting the sim right. randomisation across latency and servo stiffness + noisy ball observations were the key here. i only used open source models (robot arm, paddle, ball)
@jloganolson not yet. i used the open-source mujoco model an did some servo stiffness and latency randomisation. that was good enough for it to work (quite) reliably. it is true that it’s not the most complex problem tho
@chetan_ thanks! i did barely any classic sysid. i used the open-source so-101 arm mujoco model and randomised servo stiffness and latency per episode + trained on noisy ball estimates instead of ground truth. the hardest transfer isssues were perception and a level paddle.