ATEC2026 Online Competition: Post-Competition Review, Public Solutions, and Reproduction Entry
This directory serves as the unified entry for ATEC2026 in Chapter 15 of Every Embodied, containing task definitions, L0 task practices, Task B garbage collection review, Task D public solution analysis, Task E desktop organization reproduction, lightweight code, and public resource descriptions.
This is not a championship solution, and it does not guarantee any online scores. The tutorial breaks down a real embodied AI competition into readable, verifiable, and reproducible engineering materials, and clearly distinguishes between official evidence, local debugging evidence, and results reported by external competitors.
Directory Structure
text15-Challenge竞赛/ATEC2026/
├── README.md
├── L0-机器人徒步/ # Task A / B2Piper locomotion 复现与评估
├── L0-桌面整理TaskE/ # Task E / Piper 桌面抓取与 ACT 复盘
├── TaskB-B2Piper/ # Task B / B2-Piper 垃圾收集专题
├── docs/ # Task A/B/D/E 共性边界与发布文档
├── code/ # 轻量 adapter、参考仓库审计和检查脚本
└── references/ # 外部公开方案、固定 commit 与许可证索引
Task Overview
| Task | Robot/Ability | Entry Point of This Directory | Public Conclusion Boundaries |
|---|---|---|---|
| Task A / L0 | B2Piper Off-road Hiking | L0-机器人徒步/ |
Focus is on the official baseline, local evaluation, trajectory tracking, and checkpoint selection |
| Task B | B2-Piper Garbage Collection | TaskB-B2Piper/ |
Public complete debugging and observation boundaries; no fully verified final policy |
| Task D | Box Pushing Over Obstacles | docs/03_公开方案与Logic-TARS.md |
Focus on the public Logic-TARS state machine, LiDAR, and action adaptation |
| Task E / L0 | Piper Desk Organization | L0-桌面整理TaskE/ |
The best online result for ACT seed1 best is [15.00], not a perfect solution |
Recommended Reading
- Task definition and observation boundaries
- Problem list and solutions
- Public solution and Logic-TARS interpretation
- Reproduction, HF, and GitHub release
- Archiving reconciliation and local cleanup
- Task B special archive
Public Resources
- Official simulation repository: atecup/ATEC2026_Simulation_Challenge
- B2Piper L0 model package: Datawhale/atec2026-b2piper-l0
- Task B public reproduction archive: Datawhale/atec2026-task-b-reproducibility
- Task E model: Datawhale/atec2026-task-e-act-seed1-best
- Task E data and logs: Datawhale/atec2026-task-e-reproducibility
Reproduction Agreement
- GitHub only stores tutorials, lightweight code, experimental boundaries, fixed external commits, and masked Workspace Memory.
- Large datasets, model weights, official source code snapshots, videos, and logs are recorded on Hugging Face according to licenses and SHA-256.
- Complete simulation still requires matching Isaac Sim/Isaac Lab, GPU, Python dependencies, official runner, and evaluator; GitHub tutorials cannot be used as a one-click environment.
- Environment cameras, truth values, oracle prefix, script grippers, and phase state machines can be used for diagnosis, but protocol boundaries must be indicated; they cannot be used as formal policy inputs or complete success evidence.
- Raw training caches, private logs, tokens, server credentials, and unauthorized SDKs do not enter the public repository.