An experimental demo is available at: http://digitalopus.ca/_uploads/PhysicsCharacterController.zip
This version uses very large 512×3 layer models. The results are good, but I can realistically only run 1-2 characters per scene max.
Much of the past few months has been spent reducing the model size. I now use 256×3 models with fewer observations. These models are 10x faster. I should be able to get up to 10 characters per scene.
Also worked on a different approach of using an IK layer between the physics rig and the model. The model sees the rig limbs as extendable rods with a foot/hand on the end. It has no knowledge of the knee. The knees and elbows are managed by the IK layer.
- Faster to train
- Runs faster
- Smaller number of inputs and outputs
Unfortunately the gait is funny. I haven’t managed to train one of these with a nice gait yet. However the models are compact and efficient. I think these might be good for NPC enemies.
Also it might be possible to use the same model with a wide variety of rigs with differing numbers of leg joints. Much experimenting to do.
I also worked on running walking on moving platforms
