Embedded Files
Videos on solving tasks from:
Videos on solving tasks from:
1- Real Robot Using Raw Pixels
1- Real Robot Using Raw Pixels
2- Parkour Suite
2- Parkour Suite
3- DeepMind Control suite
3- DeepMind Control suite
4- OpenAI Gym
4- OpenAI Gym
In order to solve all these tasks we use:
In order to solve all these tasks we use:
1- One single hyperparameter set
1- One single hyperparameter set
2- Limited compute: 1 GPU for learning and 1 CPU/Robot for data generation (1 actor).
2- Limited compute: 1 GPU for learning and 1 CPU/Robot for data generation (1 actor).
Task specifications:
Task specifications:
Action dimensions range: 1 (cartpole) to 56 (CMU humanoid).
Action dimensions range: 1 (cartpole) to 56 (CMU humanoid).
State dimensions range: 5 (cartpole) to 534 (Parkour humanoid).
State dimensions range: 5 (cartpole) to 534 (Parkour humanoid).
Reward range: 50 (Parkour 2d) to 400,000 (OpenAI Gym humanoid-standup).
Reward range: 50 (Parkour 2d) to 400,000 (OpenAI Gym humanoid-standup).
Real Robot Learning From Raw Pixels:
Real Robot Learning From Raw Pixels:
Setup: 1 robot for generating data,
Setup: 1 robot for generating data,
5 action dimensions and pixels / proprioception as state
5 action dimensions and pixels / proprioception as state
Successfully learned from scratch within 100 hours of interaction to reliably lift the object using raw pixels
Successfully learned from scratch within 100 hours of interaction to reliably lift the object using raw pixels
MPO3_med.mp4Parkour Walker:
Parkour Walker:
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
6 action dimensions and 120 state dimensions
6 action dimensions and 120 state dimensions
Less than one day of training
Less than one day of training
parkour2dbest.mp4Parkour Humanoid Gaps:
Parkour Humanoid Gaps:
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
22 action dimensions and 539 state dimensions
22 action dimensions and 539 state dimensions
Less than two days of training
Less than two days of training
rtm_gaps.mp4Parkour Humanoid Walls:
Parkour Humanoid Walls:
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
22 action dimensions and 539 state dimensions
22 action dimensions and 539 state dimensions
Less than two days of training
Less than two days of training
rtm_walls.mp4rtm_walls2.mp4Humanoid CMU-Stand from DM-Control Suite:
Humanoid CMU-Stand from DM-Control Suite:
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
Setup: 1 GPU for learning and 1 CPU (actor) for generating data
56 action dimensions
56 action dimensions
policy_3060105_cam_0_video_30400.mp4Other Control Suite Tasks
Other Control Suite Tasks
humanoid_stand.mp4Humanoid-Stand
Humanoid-Stand
humanoid_run.mp4Humanoid-Run
Humanoid-Run
acrobat.mp4Double Inverted Pendulum
Double Inverted Pendulum
policy_3645919_cam_0_video_71000.mp4CartPole- Two Poles
CartPole- Two Poles
rtm_three_poles3.mp4CartPole- Three Poles
CartPole- Three Poles
rtm_bringball.mp4Manipulator
Manipulator
cheetah.mp4Cheetah
Cheetah
fastest_walker.mp4Walker
Walker
policy_1758926_cam_0_video_18600.mp4Swimmer- 15 Links
Swimmer- 15 Links
policy_720317_cam_0_video_13200.mp4Finger-Turn
Finger-Turn
policy_336392_cam_0_video_5800.mp4Finger-Spin
Finger-Spin
policy_319663_cam_0_video_5600.mp4Pendulum
Pendulum
policy_2061996_cam_0_video_26600.mp4Fish
Fish
policy_732863_cam_0_video_8800.mp4Swimmer- 6 Links
Swimmer- 6 Links
hopper4.mp4Hopper
Hopper
policy_302064_cam_0_video_6200.mp4Reacher
Reacher
OpenAI Gym Tasks
OpenAI Gym Tasks
stand_up_gym.mp4Humanoid Stand
Humanoid Stand
humanoid_run_gym.mp4Humanoid Run
Humanoid Run
Ant_Gym.mp4Ant
Ant
cheetah_gym.mp4Cheetah
Cheetah
hopper_gym.mp4Hopper
Hopper
walker2d_gym.mp4Walker2d
Walker2d
Google Sites
Report abuse