SolusOps/multilingual-vla-rollouts

Hindi and Hinglish Evaluation of StarVLA

We contribute a matched Hindi and Roman Hinglish robot-execution evaluation against an English baseline, using the frozen StarVLA Qwen3-VL + OFT policy in LIBERO-Goal.

The evaluated policy is StarVLA/Qwen3-VL-OFT-LIBERO-4in1, built on Qwen3-VL-4B-Instruct with an OFT action head. The contribution is the multilingual evaluation data and recorded instruction-sensitivity experiments; the underlying policy is StarVLA's released model.

Start here: watch a language condition

Choose multilingual_evaluation in Subset, then choose english, hindi, or hinglish in Split. Each split contains 200 matched episodes across the same ten tasks and initial-state indices 20–39. The viewer begins with video → instruction → language → task → outcome → model, followed by experiment details. IDs, filenames, hashes and media properties are at the end. Non-applicable all-null columns are omitted from individual views.

Split	Recorded task successes	Episodes	Task success
english	195	200	97.5%
hindi	58	200	29.0%
hinglish	113	200	56.5%

These 600 language-condition outcomes compare instructions under the same adapted policy and evaluation protocol. The 200 English episodes are reused from the 500-episode historical English baseline. They are not an additional independent baseline run.

What this release contributes

• Hindi and Roman Hinglish robot execution: exact instructions, paired task/state IDs, original videos and recorded native LIBERO outcomes, alongside English controls.
• French lexical exploration: all 59 wording-intervention trials, six scene-grounded French smoke episodes, and ten French-English instruction-structure episodes, organized as separate French splits.
• Instruction sensitivity and behavioral controls: the full 20-episode structure investigation, smoke tests, and six greeting/pickup controls. Behavioral controls have no success label.
• Pretrained Qwen3-VL understanding evidence: separate LIBERO image-conditioned questions, text-only questions and selected visual probes. These are VLM item sets, not robot episodes.

The release contains 1,041 original episode videos, stored once without transcoding or recompression. The viewer indexes reuse these video files. The research notebooks and analysis package explain the study. The separate Hinglish-IndicVisionBench-Benchmark contains the aligned IndicVisionBench questions and Qwen3-VL outputs.

Other subsets

Subset	Split / contents
english_baseline	500 historical English episodes, including the reused matched 200
french_explorations	lexical_trials: 59; instruction_structure: 10; scene_smoke: 6
instruction_structure	All 20 Hinglish/French-English structure episodes
smoke_tests	40 Hindi/Hinglish smoke episodes and six French smoke episodes
behavior_controls	Six greeting/pickup observations; outcome is Not scored and no success fields are supplied
excluded_fixture_pilot	Ten earlier episodes with invalid fixture pairing, kept separate from the matched evaluation
task_rollouts	Complete 1,035-row task-scene archive, organized by cohort
libero_scene_understanding	30 image-conditioned items with matched English/Hindi/Hinglish inputs and saved Qwen3-VL responses
libero_instruction_understanding	30 text-only matched items
libero_qualitative_probes	Nine selected numbered-marker/scene-probe records with exact camera inputs and outputs

The French subset includes its recorded English control within the 59 lexical trials. Subsets intentionally overlap; sum original episode IDs, not viewer subset sizes, to count recordings.

Reading outcomes correctly

For task evaluations, Success / Failure is the recorded native LIBERO goal result. Structure episodes instead show Criterion met / Criterion not met, with the recorded criterion retained: keywords count either monitored destination, while directed prompts count the requested destination first. These are different observations and remain separate from benchmark task success. Greeting/pickup controls are not scored. The invalid-fixture pilot is excluded from the reported matched metrics. Completed episodes from the stopped French smoke run are retained; no unfinished outcomes are invented.

Model and evaluation configuration

Component	Recorded setting
Framework	StarVLA
Adapted policy	StarVLA/Qwen3-VL-OFT-LIBERO-4in1
Policy revision / checkpoint	39177146f142f8805d86fcc29d2a17c2228e6073 / steps_50000_pytorch_model.pt
Vision-language backbone	Qwen/Qwen3-VL-4B-Instruct
Backbone revision	ebb281ec70b05090aa6165b016eac8ec08e71b17
Action head	OFT; action chunk 8
Environment	LIBERO-Goal; Franka; external and wrist observations at 224×224
Matched evaluator seed	7
Historical policy-server seed	Not recorded
Weight updates during language interventions	None

The LIBERO understanding item sets use the original pretrained Qwen3-VL-4B-Instruct, not the adapted policy. English historical flags come from preserved recording filenames. French exploratory conditions and the fixture pilot retain their own protocol labels. The release contains recorded video observations and evaluation metadata; action/state arrays are not supplied.

Use

from datasets import load_dataset, Video
repo = "SolusOps/multilingual-vla-rollouts"
hindi = load_dataset(repo, "multilingual_evaluation", split="hindi", token=True)
english = load_dataset(repo, "multilingual_evaluation", split="english", token=True)
hinglish = load_dataset(repo, "multilingual_evaluation", split="hinglish", token=True)
french = load_dataset(repo, "french_explorations", split="lexical_trials", token=True)
# For metadata-only access:
hindi_metadata = hindi.cast_column("video", Video(decode=False))

Access follows the repository's existing gate. The viewer provides the same data without running models or simulation.