Page 104 - Read Online

Zhang et al. Intell Robot 2022;2(3):275­97
Transformer
Quadruped
ANYmal-C
Humanoid
ANYmal
Gym
MuJoCo
RaiSim
Issac
Smooth-
Mo-
Quaternions,
CoM,
End-Effector
Foot
Joint
Velocity
Joint
Motion,
and
Tracking
Angular
Constraint,
Collisions,
Position.
Pose,
Slip.
Body
Velocities,
Body
Discrimination
Torque,
and
and
Clearance,
Velocities,
Imitation:
Positions,
Tracking,
and
Velocity
Linear
ness,
Joint
tion
Low-
Phase Offset, Joint Po-
Latent
Joint
Positions.
1)
Target.
High-Level:
Command.
Table
Positions.
Desired
sition
sired
to
(Supplement
Air-
Mo-
Ve-
States,
BaseStateandVelocities, Gravity,
Velocity
Phase,
and
Latent
Actions.
and
Velocity,
and
Forces,
Contact
CPG
Motion,
States
Goal
Pose
Previous
publications
History,
and
External
Joint
Gravity,
Base
Samples,
Positions.
and
Positions
and
Command,
Command,
States
Joint
Planning: Velocities, Base Planning: Joint States, State, Base Coordi- Planning: Stability. Slip, Foot Torque, Map. Elevation Velocity, Goal Adaption: nates. B, ANYmal Deviation, State Adaption: Goal, Feet State, Base Adaption: RaiSim Recov- Torques. Joint C ANYmal State Recovery: State. Robot Recov- Map. Elevation Torques, Joint Desired ery: Motion Foot (Planning), Space Posi- Goal Position, Joint Positions. Smoothness. Joint), (Foot, Velocity. and tions Pose, Base Tracking, Velocity JointAnglesandVelocities,Grav- Mini MIT Joint Desired IsaacGym Limits, Joint Self-Col
entations
Friction,
locities,
Height
Wheel
time.
Joint
tion,
Base
rain
gles
and
ery:
age
about
ity.
TD3, PPO PPO PPO PPO PPO V-MPO, PPO SAC information
SAC, GCPO [28] MO-VMPO,
More Gap
data- terrain. learned prior- for so- using and lever- states loco- con- tra- 2. Reality
and controller framework generation functions motion allow to switchable locomotion exteroceptive perception. locomotion robots human method for approach foot Table to
model-based drivenapproachforquadrupedal uneven over Cheetah agility. record training policy fast reward rewards captures. approach discretely integrating proprioceptive reusable legged of knowledge movement. RL proprioceptive observations control. RL-based evolutionary generator. Solution
unified locomotion Mini MIT achieving robotic achieving parallelism. Substituting stylish motion adversarial RL multiple, quadrupedal Learning real for animal end-to-end both visual motion novel an taining jectory
A A A via with from An based styles. A lution and skills prior An aging and A
IEEE Transactions Robotics 2022 2022 RSS 2022 CoRL 2022 ArXiv 2022 ArXiv Science 2022 Robotics 2022 ArXiv 2022 ICLR IEEE Robotics 2022 Autom
on
Others
Loco- Con- Using Ad- locomo- the in Learn- Movement Animal quadrupedal cross- Trajectory for
Legged Optimal Reinforcement Minutes DRL [35] Multiple RL [56] in perceptive robots Repurpose: and with Approach Locomotion [62]
Terrain-Aware and RL using via Locomotion in Walk to Parallel AdversarialMotionPriorsMakeGood Substitutes for Complex Reward Func- through Skills Priors Motion robust quadrupedal and Robot Reusable Human From vision-guided end-to-end transformers [13] Evolutionary General A
RLOC: motion trol [42] Rapid Learning [36] Learning Massively tions [91] Advanced versarial Learning for tion wild [8] Imitate ing Skills Behaviors [12] Learning locomotion modal with RL Generator: Quadrupedal Publication