MyoLeg
A lower-limb musculoskeletal model in MyoSuite containing 80 muscles. It is controlled through activation signals, and its dynamics are solved by MuJoCo.
Timeline (last 90 days)
A game animator used MyoSuite's MyoLeg model (80 leg muscles) in MuJoCo/MuJoCo Warp to train an NPC physical walking policy with the PPO algorithm (calling RSL-RL through mjlab) on a laptop GPU.
In the first attempt, after about 20 minutes of training the policy learned to walk, but peak foot force was about 4.6 times body weight and step frequency was about 230 steps per minute; with mean actions (no noise) it fell every time.
In the second attempt, after switching to sigmoid muscle commands and increasing the impact cost, foot force dropped to about 2 times body weight and the gait was smoother, but with noise turned off the policy still fell within about 1.3 seconds.
In the third attempt, after slowly reducing the action noise standard deviation (finally to 0.05), the policy could walk stably with mean actions: zero falls in 32 tests, step frequency of about 150 steps per minute, stride length of 0.44 meters, and foot force of about 1.5 times body weight.
The policy trained in the third attempt can follow speed and direction commands, start and stop (in most cases), and drive a virtual human in UE5 in real time.
The current policy has a gait asymmetry problem: stopping after a 90-degree left turn causes a fall, while a right turn does not.
MyoLeg is MyoSuite's lower-limb model, containing 80 muscles (corresponding to an 80-dimensional continuous action space), with each muscle controlled by an activation signal and dynamics solved by the MuJoCo physics engine.
All Facts (7)
A game animator used MyoSuite's MyoLeg model (80 leg muscles) in MuJoCo/MuJoCo Warp to train an NPC physical walking policy with the PPO algorithm (calling RSL-RL through mjlab) on a laptop GPU.
50%UnverifiedIn the first attempt, after about 20 minutes of training the policy learned to walk, but peak foot force was about 4.6 times body weight and step frequency was about 230 steps per minute; with mean actions (no noise) it fell every time.
50%UnverifiedIn the second attempt, after switching to sigmoid muscle commands and increasing the impact cost, foot force dropped to about 2 times body weight and the gait was smoother, but with noise turned off the policy still fell within about 1.3 seconds.
50%UnverifiedIn the third attempt, after slowly reducing the action noise standard deviation (finally to 0.05), the policy could walk stably with mean actions: zero falls in 32 tests, step frequency of about 150 steps per minute, stride length of 0.44 meters, and foot force of about 1.5 times body weight.
50%UnverifiedThe policy trained in the third attempt can follow speed and direction commands, start and stop (in most cases), and drive a virtual human in UE5 in real time.
50%UnverifiedThe current policy has a gait asymmetry problem: stopping after a 90-degree left turn causes a fall, while a right turn does not.
50%UnverifiedMyoLeg is MyoSuite's lower-limb model, containing 80 muscles (corresponding to an 80-dimensional continuous action space), with each muscle controlled by an activation signal and dynamics solved by the MuJoCo physics engine.
50%