README.md

metadata

tags:
  - sac
  - deep-reinforcement-learning
  - reinforcement-learning
  - teach-my-agent-parkour

Deep RL Agent Playing TeachMyAgent's parkour.

You can find more info about TeachMyAgent here.

Results of our benchmark can be found in our paper.

You can test this policy here.

This policy was not part of TeachMyAgent's benchmark. It was trained on the easy task space of the Parkour environment with water removed.

Results

Percentage of mastered tasks (i.e. reward >= 230) after 20 millions steps on the Parkour track.

Results shown are averages over 16 seeds along with the standard deviation for each morphology as well as the aggregation of the 48 seeds in the Overall column.

We highlight the best results in bold.

Algorithm	BipedalWalker	Fish	Climber	Overall
Random	27.25 (± 10.7)	23.6 (± 21.3)	0.0 (± 0.0)	16.9 (± 18.3)
ADR	14.7 (± 19.4)	5.3 (± 20.6)	0.0 (± 0.0)	6.7 (± 17.4)
ALP-GMM	42.7 (± 11.2)	36.1 (± 28.5)	0.4 (± 1.2)	26.4 (± 25.7)
Covar-GMM	35.7 (± 15.9)	29.9 (± 27.9)	0.5 (± 1.9)	22.1 (± 24.2)
GoalGAN	25.4 (± 24.7)	34.7 ± 37.0)	0.8 (± 2.7)	20.3 (± 29.5)
RIAC	31.2 (± 8.2)	37.4 (± 25.4)	0.4 (± 1.4)	23.0 (± 22.4)
SPDL	30.6 (± 22.8)	9.0 (± 24.2)	1.0 (± 3.4)	13.5 (± 23.0)
Setter-Solver	28.75 (± 20.7)	5.1 (± 7.6)	0.0 (± 0.0)	11.3 (± 17.9)

Hyperparameters

{'student': 'SAC'
'environment': 'parkour (easy + no water)'
'training_steps': 20000000
'n_evaluation_tasks': 100
'teacher': 'ALP-GMM'
'morphology': 'climbing_profile_chimpanzee'}