File size: 2,466 Bytes
00d189f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
---
pretty_name: AuditEmbodied
tags:
- robotics
- simulation
- evaluation
- reproducibility
---

# AuditEmbodied

AuditEmbodied contains the recorded simulation trajectories and offline evaluation inputs for *Seeing What Still Holds: How Waiting Control Changes Cooperative Recovery*. The experiments vary holding references and waiting durations during recovery from bimanual manipulation faults. The archive includes cooperative-lifting and handover trajectories and fresh evaluation rollouts from released Transport policies. [Paper](https://arxiv.org/abs/2611.22368) · [Source code](https://github.com/puppet0x3f/AuditEmbodied)

## Reproduction archive

`AuditEmbodied-review.zip` (1.90 GiB) contains 9,436 final-cohort records, including 9,400 simulator executions and 36 pre-fault initialization failures. The offline scripts regenerate 36 tables. The archive includes recorded trajectories and contact samples in `data/`, frozen lightweight evaluation models in `models/`, comparison tables in `reference/`, and the evaluators and simulator examples.

The optional Transport runners obtain the released Diffusion Policy CNN and LSTM–GMM checkpoints from their original distribution. [Checkpoint instructions](https://github.com/puppet0x3f/AuditEmbodied/blob/main/reproducibility/policy_example/public/README.md)

## Recompute the results

Download [AuditEmbodied-review.zip](https://huggingface.co/datasets/Puppeter/AuditEmbodied/resolve/main/AuditEmbodied-review.zip). Use Python 3.10–3.12 to run the recorded-trajectory evaluators on CPU.

```bash
unzip AuditEmbodied-review.zip
cd AuditEmbodied-review
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python -I recompute.py --out results/base
python -I recompute_followup.py --base-results results/base --out results/followup
python -I recompute_mechanism_geometry.py --out results/mechanism-geometry
```

Use fresh output directories for each run. The archive README describes the cohorts, sample counts, and corresponding tables.

## Citation

```bibtex
@misc{huang2026seeing,
  title         = {Seeing What Still Holds: How Waiting Control Changes Cooperative Recovery},
  author        = {Huang, Kangjie and Xin, Yucheng and Wang, Xinchun and Liu, Dong and Fan, Yan and Liu, Shuming},
  year          = {2026},
  eprint        = {2611.22368},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2611.22368},
}
```