ATEC2026 Task B: B2-Piper Garbage Collection

This is a Chinese tutorial and problem analysis guide for reproductioners. Task B is not as simple as “getting the object’s coordinates and then moving near them”; a complete closed loop should include perception, positioning, approach, control of the robotic arm and gripper, confirmation of successful grasping, handling, placement of the trash bin, and verification of the evaluation status.

First, look at the conclusion

Question	Conclusion
Is this a robot dog collecting trash?	Yes. The official task name is Garbage Collection, and the platform is B2-Piper.
Is there a global environment camera that can be directly read?	There is no default global environment camera in the official public observation contract. Head RGB-D, end RGB-D, LiDAR, and proprioception are the main sources of observation.
Can object coordinates be directly read?	During formal operation, it cannot rely on the true pose of the object, random seeds, or internal object lists in simulation. The target position should be estimated from legitimate sensor observations.
Is reaching the target pose considered complete?	No. Approaching, touching, gripping, lifting, and placing in the trash can all require separate verification. Being close enough may only be a contact agent score and cannot be considered a successful physical grasping.
Can other items be transported?	It can transport state machines, visual servo, whole-body coordination, calibration, and diagnostic methods; but it cannot directly transport robot assets, joint orders, camera parameters, or reward assumptions.
Is there a complete public Task B champion implementation that can be verified currently?	No such complete public implementation with official verifiable final scores and weights is found in this archive. The closest to a complete process is the SLAM public fork; see for external solution comparison .

Public Code

File	Purpose
[`code/inspect_task_b_observation.py`](…/…/…/…/15-Challenge competition/ATEC2026/TaskB-B2Piper/code/inspect_task_b_observation.py)	Does not start simulation; checks the keys, types, and shapes of masked observation samples
[`code/build_public_manifest.py`](…/…/…/…/15-Challenge competition/ATEC2026/TaskB-B2Piper/code/build_public_manifest.py)	Generates a file list with SHA-256 for the public directory
[`code/scan_public_release.sh`](…/…/…/…/15-Challenge competition/ATEC2026/TaskB-B2Piper/code/scan_public_release.sh)	Checks credentials, absolute machine paths, and large files
[`code/fetch_reference_repos.sh`](…/…/…/…/15-Challenge competition/ATEC2026/TaskB-B2Piper/code/fetch_reference_repos.sh)	Pulls external public reference repositories into a cache directory outside the repository
[`configs/observation_contract.yaml`](…/…/…/…/15-Challenge competition/ATEC2026/TaskB-B2Piper/configs/observation_contract.yaml)	Declarative configuration for the Task B observation boundaries and prohibited dependencies

These scripts are tutorials and deployment tools, and they do not mimic a complete scoring policy. When actually running the competition code, you must use the official runner, official environment, and official tester.

Quick Check

bashcd /path/to/every-embodied/15-Challenge竞赛/ATEC2026/TaskB-B2Piper
bash code/scan_public_release.sh .
python3 code/build_public_manifest.py . --output public_manifest.json

If you need to check a masked JSON observation example:

bashpython3 code/inspect_task_b_observation.py --input observation_sample.json

If you need to obtain a public reference implementation, use a directory outside the repository:

bashbash code/fetch_reference_repos.sh "${TMPDIR:-.reference-cache}"

Evaluation Evidence

The current archive prioritizes retaining verifiable boundaries, rather than using a video or a close-up shot to claim success. The experiment records can be found in [results/experiment_ledger.csv](…/…/…/…/15-Challenge competition/ATEC2026/TaskB-B2Piper/results/experiment_ledger.csv), and the evidence explanations are available in results/evidence_boundary.md.

Hugging Face

For details on the public archive, see [hf/README.md. The HF dataset repository contains this topic documents, release lists, verified official public source code snapshots, debugging media, and necessary historical materials; it does not upload private rights, private logs, or unauthorized large external files. For the complete upload mapping and local deletion list, see archive reconciliation and local cleanup .