Datasets:

nur-dev
/

multilingual-robot-grounding-causal-audit

license: cc-by-4.0
language:
  - en
  - ru
  - kk
tags:
  - robotics
  - embodied-ai
  - human-robot-interaction
  - multilingual
  - speech-recognition
  - text-to-speech
  - computer-vision
  - robot-manipulation
  - vision-language-action
  - causal-evaluation
  - semantic-grounding
  - reproducibility
  - kazakh
  - russian
  - english
pretty_name: 'Multilingual Robot Grounding: A Causal Speech–Vision–Action Audit'
size_categories:
  - n<1K

Multilingual Robot Grounding: A Causal Speech–Vision–Action Audit

Frozen manifests, per-condition traces, semantic-preservation analyses, instruction-dependence controls, counterfactual robot outcomes, and reproducibility scripts for a multilingual speech–vision–action grounding audit.

A compact multilingual (Kazakh / Russian / English) speech→ASR→parser→perception→policy→TTS pipeline is evaluated on a frozen benchmark of 60 counterfactual VLABench states (480 conditions, 960 paired closed-loop episodes), with and without a deterministic six-action safeguard. This repository contains read-only trace data and derived tables/figures only; no model weights and no secrets are included.

Provenance

Frozen results package final_results_package.json, sha256 prefix b1c42a62bc46b0b2.
Normalizer e7b53b46…; guard deterministic_compact_guard.v1; command spec 53487b00cd7a6b2a; fixtures 93181e7a6b25add8; frames 2889872014f3ff49; ASK fixtures 027dae74be94338a; episodes a626e682d055e03c; gold results 47b41f2d2dbeb9ea; policy checkpoint lerobot/smolvla_vlabench.
No large-model supervision, distillation or oracle leakage. All 960 primary episodes completed with zero technical failures. SHA256SUMS.txt fixes every file below.

Pipeline components (referenced, not redistributed here)

ASR — nur-dev/Qwen3-ASR-0.6B-kk-ru-en
Semantic parser — nur-dev/farabi-0.6B-agent-rag
TTS — nur-dev/ait-syn-0.6b-trilingual-v30
Policy — lerobot/smolvla_vlabench

Path	What
`final_results_package.json`	Single frozen package; source of every headline number.
`analyses/`	Canonical ASR×frame cross-tab, stage decomposition, exact intervals, gold-input analysis, paired study analysis.
`traces/`	Per-condition raw traces: episodes, fixtures, S0 frames, ASK fixtures, benchmark conditions.
`figures/`	The five paper figures (vector PDF).
`tables/`	The ten paper tables (LaTeX `tabular` bodies).
`native_review_worksheet.csv`	The 26 cross-script ASR recoveries pending native Kazakh/Russian review.
`SHA256SUMS.txt`	Checksums for all of the above.

Reproducing the tables and figures

The tables and figures are regenerated purely from final_results_package.json (plus the frozen instruction-dependence audit) by the paper's reproduce/ scripts:

python make_tables.py  --package final_results_package.json --e1b <audit.json> --out tables/
python make_figures.py --package final_results_package.json --e1b <audit.json> --out figures/

Scope and limitations

Simulation only; synthetic integration speech, not human speech. Object results are exploratory (2 templates); the attribute axis is primary (5 templates). Destination and relation axes are not claimed. 26 cross-script ASR-recovery labels are pending native review and can affect only those 26 condition labels, none of the execution results.

License

Released under CC-BY-4.0.