LeHome Challenge 2026 - VLMRoutedMoE Submission

Overview

Key Features

1. Dynamic Shader Warmup: An observation wrapper automatically monitors incoming simulation streams and extends the default 15-step warmup to up to 150 steps if it detects incomplete engine illumination/shader rendering (median brightness < 100).
2. Robust Multi-Scale Policies: We utilize a highly stabilized pant_long expert trained over 200,000 steps to combat complex topological deformations.
3. Elegant Proxy Fallback: To deal with variable internet access restrictions during evaluations, the codebase prioritizes a direct POST to generativelanguage.googleapis.com. If the connection times out natively, it dynamically invokes a local proxy fallback strategy.

File Structure

• vlm_routed_moe.py: Main evaluating policy Python script.
• submission_models/: Contains the four trained LeRobot checkpoints (act_top_long_100k, act_pant_long_200k, act_top_short_100k, act_pant_short_100k).

Note to Evaluators