Research Scientist / Engineer - Data & Evaluation

At Rhoda AI, we’re building the next generation of generalist intelligent robots. We own the full robotics stack from high-performance hardware and robot systems to the infrastructure and state-of-the-art foundation world models that control our robots. Our robots are designed to be generalists capable of operating in complex, real-world environments and handling long-tail edge cases, made possible by our cutting edge research and end-to-end system design. We've raised over $400M and are investing aggressively in model research, infrastructure, hardware development, and manufacturing scale-up to make generalist robotics a reality.

We're looking for Research Scientists and Research Engineers to build the data and evaluation foundations for our video action model. This team owns web-scale video data curation, annotation pipelines, and evaluation methodology — directly determining the quality of the video pretraining distribution and how clearly we can measure model progress. We hire across levels — from MTS-Staff

What You'll Do

• Design and implement scalable curation pipelines for web-scale video pretraining data: ingestion, deduplication, quality filtering, and content classification across internet-scale video corpora
• Develop video-specific annotation frameworks and quality filters — motion quality, scene diversity, action content, temporal coherence — to improve pretraining signal
• Build evaluation frameworks and benchmarks to measure causal video model capabilities: prediction quality, temporal coherence, long-horizon rollout fidelity, and downstream robot task performance
• Research and implement data selection, mixing, and weighting strategies that improve video generation quality and transfer to robotic control
• Deploy and scale vision-language models (VLMs) and video understanding models for automated annotation, filtering, and content scoring at web scale
• Collaborate closely with pre-training and post-training teams to ensure data quality and evaluation methodology drive research decisions
• Track model capability trends across training runs, catching regressions and surfacing improvements early

What We're Looking For

• Strong understanding of data-centric ML and how web video data quality affects large generative model performance
• Experience building large-scale video data pipelines: ingestion, filtering, deduplication, and quality scoring
• Familiarity with video-specific data characteristics: temporal structure, motion quality, scene diversity, and action content
• Solid ML fundamentals with hands-on experience training or evaluating large generative models
• Ability to design evaluations for video generation models that are diagnostic, reproducible, and actionable
• Staff-level candidates are expected to define technical direction and drive research strategy independently; senior/MTS candidates execute complex projects with strong fundamentals and growing scope

Nice to Have (But Not Required)

• PhD or strong research background in ML, computer vision, or a related field
• Experience with large-scale web video dataset curation (e.g., WebVid, HowTo100M, Ego4D, or similar)
• Familiarity with video generation quality metrics (FVD, perceptual quality, motion consistency)
• Experience running VLM or CLIP-style inference at scale for automated video filtering and annotation
• Prior work on evaluation methodology for video generation or world models
• Understanding of how web video data properties connect to downstream robotic action prediction
• Publication record at NeurIPS, ICML, ICLR, CVPR, or related venues

Why This Role

• The video curation and evaluation rigor you build directly determines pretraining quality and research iteration speed for the entire team
• Build the benchmark infrastructure that gives the team an honest signal of model progress toward real robot performance
• High leverage: improvements to data quality compound across every training run
• Work at the intersection of large-scale systems and generative model research with visibility across all model development