Skip to content

MarathonEnvs-v2.0.0
Compare
Choose a tag to compare
@Sohojoe Sohojoe released this 13 Jun 05:18
· 78 commits to master since this release
v2.0.0
b95d96c
This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
GPG key ID: 4AEE18F83AFDEB23
Expired
Verified
Learn about vigilant mode.

What's new in MarathonEnvs-v2.0.0

New Style Transfer Environments

  • MarathonManWalking-v0
  • MarathonManRunning-v0
  • MarathonManJazzDancing-v0
  • MarathonManMMAKick-v0
  • MarathonManPunchingBag-v0
  • MarathonManBackflip-v0
  • Plus various fixes to improve performance and make it closer to DeepMimic paper.

Guide to Working With Style Transfer

  • Guide on how to train your custom motion capture sequence.

WebGL Demo / Support for in browser

marathon-envs Gym wrapper (Preview)

  • Use marathon-envs as a OpenAI Gym environment - see documentation

ml-agents 0.14.1 support

  • Updated to work with ml-agents 0.14.1 / new inference engine

Unity 2018.4 LTS

  • Updated to use Unity 2018.4 LTS. Should work with later versions. However, sometimes Unity makes breaking physics changes.

MarathonManBackflip-v0

  • Train the agent to complete a backflip based on motion capture data
  • Merged from StyleTransfer experimental repro

MarathonMan-v0

  • Optimized for Unity3d + fixes some bugs with the DeepMind.xml version
  • Merged from StyleTransfer experimental repro
  • Replaces DeepMindHumanoid

ManathonManSparse-v0

  • Sparse version of MarathonMan.
  • Single reward is given at end of the episode.

TerrainHopperEnv-v0, TerrainWalker2dEnv-v0, TerrainAntEnv-v0, TerrainMarathonManEnv-v0

  • Random Terrain environments
  • Merged from AssaultCourse experimental repro

SpawnableEnvs (Preview)

  • Set the number of instances of an environment you want for training and inference
  • Environments are spawned from prefabs, so no need to manually duplicate
  • Supports ability to select from multiple agents in one build
  • Unique Physics Scene per Environment (makes it easier to port environments however runs slower)
  • SelectEnvToSpawn.cs - Optional menu to enable user to select from all agents in build

Scorer.cs

  • Score agent against 'goal' (for example, max distance) to distinguish rewards from goals
  • Gives mean and std-div over 100 agents

Normalized Observations (-1 to 1) and reward (0 to 1)

  • No need to use normalize flag in training. Helps with OpenAI.Baselines training

Merge CameraHelper.cs from StyleTransfer. Controls are

  • 1, 2, 3 - Slow-mo modes
  • arrow keys or w-a-s-d rotate around agent
  • q-e zoom in / out

Default hyperparams are now closer to OpenAI.Baselines

  • (1m steps for hopper, walker, ant, 10m for humanoid)

Training speed improvements - All feet detect distance from floor

Assets 4
Loading