| Note: Please use Isaacgym v3 rather than v4 to aviod robot control issues.
- Fast: Fully GPU-Based Pipeline (isaac support, fast buffer indexing, fast batch relabeling)
- Designed for Sparse Reward (HER)
- Easy to use: single config file, multi-algorithm support(Red-Q, SAC, TD3, DDPG, PPO), resume from cloud, etc.
python run.py
'''debug'''
# do the environment check
python run.py +exp=1ho_debug
'''train'''
# run main experinment
python run.py +corl=nho
'''eval'''
python run.py +corl=2ho_render
'''env'''
# render env with handwriting policy
python envs/franka_cube.py -e
# render env with random policy
python envs/franka_cube.py -r
add these line to config file to run automatic curriculum:
curri:
table_gap: # env params to change
now: 0
end: 0.2
step: 0.05
bar: -0.2 # reward bar to change curriculum params
- HER relabel lead to fluctuations in success rate
- check relabel boundary case
- check relabel reward, mask
- check trajectory buffer
- check
- Add SAC
- Add PPO
- Add RedQ
- Add HER
- add info (env_step, traj_idx) to HER
- add trajectory extra info (traj_len, ag_pool) to HER
- Toy Env
- reach
- pnp
- handover
- Eval Func (in agent)
- Normalizer
- logger
- check buffer function for multi done collect
- env reset function
- Add Hydra
- sub task evaluation
- hand write normalizer data check
- Add Attention Dense Net
- Resume run from cloud
- Add Isaac Env (fast auto reset)
- PNP
- Handover
- Multi Robot Environment
- Isaac render + viewer
- Render function
- merge all buffer together (fast indexing)
-
get_env_params
,obs_parser
,info_parser
function - curriculum learning
- add ray tune (wandb)
- update according to collected steps
- merge buffer into agent
- fix relabel for to generate to left index
- add vec transitions at a time
- resume from remote
- update env info dim automatically
- classified log info
see requirements.txt
.