Skip to content
Research into locomotion style transfer with Active Ragdolls (using MarathonEnvs +ml_agents)
Branch: master
Clone or download
Latest commit 6a3399c Jul 7, 2019

README.md

ActiveRagdollStyleTransfer

Research into using mocap (and longer term video) as style reference for training Active Ragdolls / locomotion for Video Games

(using Unity ML_Agents + MarathonEnvs)


Goals


Using this repro

  • Make sure you are using a compatable version of Unity (tested with 2018.4 LTS and 2019.1)

  • To run trained models, make sure you: add TensorFlowSharp to Unity

  • To try different moves: Replace reference MoCap in animation tree and select the right ML-Agent trained model

  • To re-train:

    • Set the LearnFromMocapBrain to External SetBrainType.png
    • Build the project
    • From the root path, invoke the python script like this: mlagents-learn config\style_transfer_config.yaml --train --env="\b\StyleTransfer002\Unity Environment.exe" --run-id=StyleTransfer002-145 where "\b\StyleTransfer002\Unity Environment.exe" points to the built project and StyleTransfer002-145 is the unique name for this run. (Note: use / if on MacOS/Linux)
    • See the ML-Agents documentation for more details on using ML-Agents
  • Post an Issue if you are still stuck


Contributors


Download builds : Releases


StyleTransfer002

Backflip (002.144-128m steps)
StyleTransfer002.128
Running (002.114) Walking (002.113)
StyleTransfer002.114 StyleTransfer002.113
  • Model: MarathonMan (modified MarathonEnv.DeepMindHumanoid)
  • Animation: Runningv2, Walking, Backflip
  • Hypostheis: Implement basic style transfer from mo-cap using MarathonEnv model
  • Outcome: Is now training on Backflip
    • Initial was able to train walking but not running (16m steps / 3.2m observations)
    • Through tweaking model was able to train running (32m steps / 6.4m observations)
    • Was struggling to train backflip but looks like I need to train for longer (current example is 48m steps / 9.6m observations)
    • Was able to train Backflip after updating to Unity 2018.3 beta - looks like updates to PhyX engine improve stability
  • References:
  • Notes:
    • Needed to make lots of modifications to model to improve training performance
    • Added sensors to feet improved trainging
    • Tweaking joints improved training
    • Training time was = ~7h for 16m steps (3.2m observations) TODO: check assumptions
    • New Training time is + 2x
    • ... Optimization: Hack to Academy to have 4 physics only steps per ml-step
    • ... Optimization: Train with 64 agents
    • ... also found training in headless mode --no-graphics helped
    • Updated to Unity 2018.3 Beta for PhysX improvements
    • see RawNotes.002 for details on each experiment

StyleTransfer001

StyleTransfer001

  • Model: U_Character_REFAvatar
  • Animation: HumanoidWalk
  • Hypostheis: Implement basic style transfer from mo-cap
  • Outcome: FAIL
    • U_Character_REFAvatar + HumanoidWalk has an issue whereby the feet collide. The RL does get learn to avoid - but it feels that this is slowing it down
  • References:
    • Insperation: [DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills arXiv:1804.02717 [cs.GR]](https://arxiv.org/abs/1804.02717
  • Raw Notes:
    • Aug 27 2018: Migrate to new repro and tidy up code so to make open source
You can’t perform that action at this time.