video video 30.1 58.6 | label class label 2
classes |
|---|---|
0hrobot20dof230mm | |
0hrobot20dof230mm | |
0hrobot20dof230mm | |
0hrobot20dof230mm | |
0hrobot20dof230mm | |
1hrobot22dof240mm | |
1hrobot22dof240mm |
- Videos
hrobot20dof230mm/hrobot20dof230mm_walk.mp4hrobot20dof230mm/hrobot20dof230mm_getup.mp4hrobot20dof230mm/hrobot20dof230mm_rough.mp4hrobot20dof230mm/hrobot20dof230mm_slopes.mp4hrobot20dof230mm/hrobot20dof230mm_steps.mp4hrobot22dof240mm/hrobot22dof240mm_walk.mp4hrobot22dof240mm/hrobot22dof240mm_getup.mp4
- License
Hrobot media
Videos of the Hrobot small humanoid robots in simulation.
| Robot | Description |
|---|---|
| Hrobot20dof230mm | 230 mm, 20 DoF, 20 Dynamixel XL330 servos (five joints per leg, no ankle roll); current design |
| Hrobot22dof240mm | 240 mm, 22 DoF, 22 Dynamixel XL330 servos |
| Hrobot22dof540mm | 540 mm, 22 DoF |
Videos
hrobot20dof230mm/hrobot20dof230mm_walk.mp4
Hrobot20dof230mm, the current 20-DoF design, walking with its learned policy in MuJoCo physics (not a kinematic animation), in real time, 51 s, with the same command plan as the 22-DoF video: standing, walking forward up to 0.25 m/s, turning, walking sideways and backward, and a side push.
hrobot20dof230mm/hrobot20dof230mm_getup.mp4
Hrobot20dof230mm getting up with its learned policy in MuJoCo physics (not a kinematic animation), 48 s. Each start, lying face down, face up and on each side, is shown at rest after a 1 s settle (servos hold the pose, as after a real fall), then the get-up in real time and four times slower. Unlike the first, acrobatic get-up of Hrobot22dof240mm, this policy was trained with penalties on fast trunk rotation, flight, servo torque near stall and head contact, introduced gradually during training: the robot pushes up onto hands and feet and rises from a crouch, without turning over its head and without leaving the ground.
hrobot20dof230mm/hrobot20dof230mm_rough.mp4
Hrobot20dof230mm walking over rough terrain with its rough-terrain policy (the flat walking policy fine-tuned for 1000 iterations on rough terrain) in MuJoCo physics (not a kinematic animation), in real time, 48 s. The policy is blind: it has no terrain sensing. A straight course has bumps, stairs up and down, and slopes, first at half and then at the full training difficulty (bumps 6 mm, stairs 12 mm, slope 6.8 degrees). The command is 0.2 m/s forward, with small steering corrections that keep the robot on the course. The robot crosses the bumps and slopes without a fall but cannot climb a single 6 mm step: it steps in place in front of it and is then put back behind the stairs.
hrobot20dof230mm/hrobot20dof230mm_slopes.mp4
The same rough-terrain policy on ever steeper slopes (6.8, 10, 13, 16, 20 and 25 degrees; the training maximum was 6.8), each up a 0.4 m ramp, over a 0.3 m platform and down again, in MuJoCo physics (not a kinematic animation), in real time, 59 s. The command is 0.2 m/s forward. The robot walks up slopes of up to 10 degrees; from 13 degrees it steps in place at the foot of the ramp (25 degrees: it falls), and it is then put on the top platform to test going down. It walks down slopes of up to 20 degrees, though faster than commanded, and falls at the bottom of the 25-degree slope.
hrobot20dof230mm/hrobot20dof230mm_steps.mp4
The same rough-terrain policy on stairs of 2, 4, 5 and 6 mm steps (four risers up, a 0.4 m platform, four risers down, 0.10 m treads), in MuJoCo physics (not a kinematic animation), in real time, 40 s. The command is 0.2 m/s forward. The blind policy climbs and descends the 2, 4 and 5 mm stairs; at the first 6 mm step it steps in place, since it lifts its feet only about 7 mm, and it is then put on the top platform, from where it walks down the 6 mm stairs.
hrobot22dof240mm/hrobot22dof240mm_walk.mp4
Hrobot22dof240mm walking with a learned policy in MuJoCo physics (not a kinematic animation), in real time, 51 s. The policy was trained with mjlab (MuJoCo Warp, PPO) on flat ground and runs at 50 Hz; physics runs at 200 Hz with a DC model of the XL330 servo. The video shows standing, walking forward up to 0.25 m/s, turning, walking sideways and backward, and recovering from a side push. The commanded and measured velocities are shown on screen.
hrobot22dof240mm/hrobot22dof240mm_getup.mp4
Hrobot22dof240mm getting up with a learned policy in MuJoCo physics (not a kinematic animation), 30 s. It starts lying face down, face up, and on each side, as in training; each case is shown in real time and then four times slower. This first get-up policy rises in about half a second with a fast, acrobatic motion (pushing off with the arms and briefly leaving the ground); a slower, gentler get-up for the real robot is planned.
License
The videos are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0). You may share and adapt them for any purpose, provided you give appropriate credit to the Hrobot project (Dejan Batanjac and Vladan Vučković).
- Downloads last month
- 81