Skip to content
MyoHubPublic

About

MyoChallenge 2026 — participant submission guide and policy examples

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

17 Commits

Folders and files

Repository files navigation

MyoChallenge 2026 · ChaseTag

Build a full-body policy that can both catch and evade. Submit one controller, then compete in a balanced title-cycle format on the MyoChallenge site.

Environment Roles Observation → action Episode
myoChallengeChaseTagFBVs-v0 Chase + evade 546 → 354 10 seconds

Access: invitation-only. An organizer must approve your team before submission. The site shows current track availability and cycle status.

ChaseTag uses one policy for both chase and evade roles

On this page: Start · Submission format · Task contract · Title cycle · Training

Start here

flowchart LR
  A[Register team] --> B[Install evalmc]
  B --> C[Export policy to ONNX]
  C --> D[Submit]
  D --> E[Track evaluation]
Loading
  1. Register at mc26.myosuite.org and save your CLI token.

  2. Install and configure the CLI:

    pip install evalmc
    evalmc host --set-host https://api.mc.myosuite.org
    evalmc login
    evalmc status
  3. Submit an ONNX model:

    evalmc submit-onnx examples/artifacts/chasetag_walker.onnx -t chasetag -p 1

    This repo includes a tested starter model at that path. It runs in both roles but is a starter policy, not a guarantee of competitive performance. Confirm the file exists and use the local export instead if you have trained a policy. ONNX input is [batch, 546]; output is [batch, 354].

  4. Follow the submission in the website’s Submissions tab. Evaluation runs asynchronously.

New to submissions? Follow the ChaseTag quickstart.

Submission format

ONNX is the only accepted submission format. Export your policy as a single ONNX inference graph and submit it with:

evalmc submit-onnx model.onnx -t chasetag -p 1
Tensor Shape Type
input obs [batch, 546] float32
output action [batch, 354] float32

Any preprocessing (normalization, role handling) must be part of the exported graph. Recurrent policies (GRU/LSTM) can add <name>_in / <name>_out state tensors, zeroed at each episode start. See the ONNX submission contract for details.

Task contract

MyoSuite 3.1+ runs ChaseTag on the myofullbody model. Your policy controls agent_0; the opponent is agent_1. Role is included in the observation’s final two values:

Role values Meaning
[1, 0] Chase: tag the opponent to win
[0, 1] Evade: avoid being tagged until time expires

Both roles appear in balanced, shuffled bouts. Do not infer your role from the opponent’s position. ChaseTag policies trained on the older MuscleMimic body may need retraining for MyoSuite 3.1+.

Start positions are randomized. In evaluation, both players' starting positions and headings vary slightly from the default MyoSuite reset in every episode. The change is mirrored, so both players get the same relative start, and each chase/evade pair of episodes shares it. The exact ranges are not published: train policies that are robust to small changes in the starting pose, not only to the default reset.

Title cycle

flowchart LR
  A[Mid-cycle bouts<br/>6 episodes] --> B[Freeze<br/>eligible entries seeded]
  B --> C[Finals<br/>single elimination]
  C --> D[Next cycle champion]
Loading
  • Mid-cycle bouts use 6 episodes: 3 chase and 3 evade. Every submission faces the fixed cycle champion; a win does not take the crown.
  • Before freeze, nominate the scored submission you want considered for finals. Without a nomination, the best eligible in-window submission is used.
  • Finals bouts use 10 episodes: 5 chase and 5 evade. The defending champion is seed #1.
  • Win Rate is episode wins divided by wins plus losses; draws are excluded. The leaderboard also reports role-specific rates and outcome counts.
  • Ties use episode wins and then earlier scoring for mid-cycle ranking.
  • A finals bout tied after 10 episodes is settled by the first of these that separates the teams:
    1. Overtime: 2 more episodes (one each way) from new start positions, with the players starting back to back.
    2. More successful tags while chasing.
    3. More escapes while evading.
    4. Fewer falls while chasing.
    5. Closest approach while chasing (smaller average closest distance to the runner).
    6. A coin flip. Seeding is never used as a tiebreaker.

Exact dates, access, and current cycle status are shown on the competition site, which is the source of truth for live operations.

Train and test

Install MyoSuite 3.1+ to run the environment locally:

python -m pip install -U 'myosuite>=3.1.0'
python -c "from myosuite import make_env; env=make_env('myoChallengeChaseTagFBVs-v0'); print(env.observation_space, env.action_space); env.close()"

See training, checkpoint, and export for the MyoSuite ChaseTag Colab workflow and ONNX export.

Help

About

MyoChallenge 2026 — participant submission guide and policy examples

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages