L0 B2Piper option-MoE Submission Code Skeleton

This directory contains the publicly submitable code skeleton for ATEC2026 Online Competition, Track 1 L0 “Robot Hiking”.

File description:

• solution.py: Official website call entry.
• solution_rl.py: B2Piper low-level policy loading, speed command schedule, manual/rough option-MoE logic.
• requirements.txt: Example is empty, indicating no additional pip packages need to be installed, and the dependency is built into the official image environment.

Note:

1. No policy.pt is included here. Please add the same directory from the official baseline, your training export results, or weights that are publicly available in competitions.
2. By default, this code loads policy.pt(https://github.com/lllyu/LLaVA/tree/master/code/ee/ee_keep_0001) from the same directory, assuming it is the TorchScript locomotion actor for 45 -> 12.
3. rough option-MoE represents an engineering attempt during post-competition analysis: retaining the protected gait prior, and using the score segment and progress gate to carefully switch between the rough and smooth segments to execute/escape commands.
4. It is not a championship solution, nor does it guarantee full marks; it is more suitable as a reference for “how to encapsulate the competition strategy logic into solution.py”.

Example of the submission directory structure:

textsolution.py
solution_rl.py
policy.pt
requirements.txt

Local smoke:

bashpython - <<'PY'
import torch
m = torch.jit.load("policy.pt", map_location="cpu")
print(m(torch.zeros(1, 45)).shape)
PY

Expected output:

texttorch.Size([1, 12])