-
OpenAI, "Learning Dexterous In-Hand Manipulation", arXiv:1808.00177, 2018. [Paper]
Parameter Scaling factor range Additive term range object dimensions uniform([0.95, 1.05]) object and robot link masses uniform([0.5, 1.5]) surface friction coefficients uniform([0.7, 1.3]) robot joint damping coefficients loguniform([0.3, 3.0]) actuator force gains (P term) loguniform([0.75, 1.5]) joint limits $N$ (0, 0.15) radgravity vector (each coordinate) $N$ (0, 0.4)$m/s^2$ -
X. B. Peng, et al., "Sim-to-Real Transfer of Robotic Control with Dynamics Randomization", ICRA, 2018. [Paper] [Blog]
Parameter Range Link Mass [0.25, 4] x default mass of each link Joint Damping [0.2, 20] x default damping of each joint Puck Mass [0.1, 0.4]$kg$ Puck Damping [0.01, 0.2]$Ns/m$ Puck Friction [0.1, 5] Table Height [0.73, 0.77]$m$ Controller Gains [0.5, 2] x default gains Action Timestep $\lambda$ [125, 1000]$s^{-1}$ -
J. Tan, et al., "Sim-to-Real: Learning Agile Locomotion For Quadruped Robots", RSS, 2018. [Paper]
Parameter Lower bound Upper bound mass 80% 120% motor friction 0$Nm$ 0.05$Nm$ inertia 50% 150% motor strength 80% 120% control step 3$ms$ 20$ms$ latency 0$ms$ 40$ms$ battery voltage 14.0$V$ 16.8$V$ contact friction 0.5 1.25 IMU bias -0.05 radian 0.05 radian IMU noise (std) 0 radian 0.05 radian -
L. Pinto, et al., "Robust Adversarial Reinforcement Learning", PMLR, 2018. [Paper]
- Robustness check using trained policy by changing mass, friction.
-
A. Rajeswaran, et al., "EPOpt: Leaerning Robust Neural Network Policies Using Model Ensembels", ICLR, 2017. [Paper]
- Robustness check using trained policy by changing mass, friction.
Hopper $\mu$ $\sigma$ low high mass 6.0 1.5 3.0 9.0 ground friction 2.0 0.25 1.5 2.5 friction damping 2.5 1.0 1.0 4.0 armature 1.0 0.25 0.5 1.5 Half-Cheetah $\mu$ $\sigma$ low high mass 6.0 1.5 3.0 9.0 ground friction 0.5 0.1 0.3 0.7 friction damping 1.5 0.5 0.5 2.5 armature 0.125 0.04 0.05 0.2