Senior, ML Engineer - VLM

Torc Robotics • Ann Arbor, MI

9d

About The Position

At Torc, we are focused solely on developing software for automated trucks to transform how the world moves freight. This team owns the dataset layer that powers end-to-end Vision-Language-Action (VLA) models. The team turns petabytes of logged multi-modal fleet data into VLM/VLA-ready datasets, including geometric annotations, semantic descriptions, action- and trajectory-grounded labels, and reasoning traces. They run a continuous data flywheel: mining long-tail and failure cases, auto-labeling at scale, validating quality, and feeding curated datasets into Torc’s end-to-end VLM/VLA model development.

Requirements

• Bachelor’s Degree in Computer Science, Robotics, Electrical Engineering, or related technical field plus competences typically acquired through 6+ years of experience; OR Master’s Degree in a related technical field plus competences typically acquired through 3+ years of experience.
• Computer Vision & Deep Learning — model training and at least two of: 2D/3D Object Detection, Tracking, Sensor Fusion, Semantic Segmentation, BEV, Depth Estimation.
• Multimodal / VLM experience — hands-on work with vision-language models, open-vocabulary or zero-shot recognition, dense captioning, or semantic embeddings / search applied to perception data.
• Model Data Curation — building targeted datasets that measurably improve downstream model performance; large-scale Parquet data processing (Databricks, Daft, Pandas, etc.).
• Distributed ML & data frameworks — PyTorch, Lightning, Ray, Spark, or equivalent for training and large-scale data processing.
• Scaled MLOps & Tooling — experiment tracking, model registry, MLflow / Weights & Biases, and ML metrics, evaluation, and quality.
• Development Tools & Eco-System (at scale) — strong Python software development, VDI and cloud-based development environments, CI systems (GitHub Actions), and Docker.

Nice To Haves

• End-to-end / VLA driving — familiarity with VLM/VLA or end-to-end driving models, trajectory and action grounding, or chain-of-causation / reasoning-trace datasets.
• Auto-labeling foundation models — experience with segmentation, open-vocabulary detectors, or VLM/LLM-driven data engines for annotation and verification.
• High-throughput model serving — vLLM, SGLang, or similar for batch auto-labeling and inference at scale.
• Semantic inference & retrieval — attribute mapping, semantic search, and vector databases (e.g., LanceDB) for automotive data.
• AV data standards & tooling — scenario-description standards such as Pegasus layers; parsing robotics formats (ROS bags, MCAP) and optimizing columnar storage (Parquet, Arrow).
• Cloud development & orchestration — Terraform and AWS managed services (S3, ECS, Lambda, DynamoDB, Step Functions, Athena); AWS HyperPod / Anyscale; inference orchestration.
• Data visualization — Foxglove, FiftyOne (51), three.js, OpenGL, or similar for dataset inspection and accessibility.
• Evaluation & research — closed-loop / open-loop evaluation frameworks (e.g., NavSim-style planning metrics); publications in top-tier CV/AI/Robotics venues (CVPR/ECCV/ICCV, NeurIPS/ICLR/ICML, CoRL).

Responsibilities

• Own the offline dataset pipeline — design, implement, test, and deploy Cloud-based pipelines that convert logged multi-sensor data into VLM/VLA training datasets, spanning geometric labels (3D/2D detection, tracking, segmentation, depth) through semantic, scenario-level, and action/trajectory-grounded annotations.
• Build VLM-assisted auto-labeling — develop open-vocabulary detection, dense captioning, semantic enrichment, and scene/scenario description generation that move beyond closed-set bounding boxes, using foundation models to scale annotation and cut manual labeling cost.
• Generate reasoning-grounded labels — produce language-grounded reasoning and chain-of-causation style annotations, temporally aligned to ego-motion and trajectories, to support VLA training and explainable driving behavior.
• Mine and curate the long tail — surface rare, difficult, and high-uncertainty scenarios, and build curated datasets that measurably improve downstream VLM/VLA model metrics rather than simply adding volume.
• Close the data flywheel — define dataset schemas, quality metrics, and validation; track auto-labeling quality against model requirements; route model failures back into re-labeling and retraining loops.
• Partner with the end-to-end model team — co-define dataset specifications with VLM/VLA model developers, own the quality bar and delivery cadence, and operationalize a continuous dataset delivery loop into their training pipelines.
• Scale on cloud infrastructure — build distributed, reproducible pipelines using columnar data formats and distributed compute, with disciplined software practices, version control, and documentation.
• Lead and mentor — serve as project lead, guide less-experienced engineers, run design reviews, set coding and annotation standards, and drive alignment across team interfaces to the rest of the organization.
• Stay current — track the latest advances in multimodal models, auto-labeling, and end-to-end autonomous driving, and translate relevant research into production data systems.

Benefits

• A competitive compensation package that includes a bonus component and stock options
• 100% paid medical, dental, and vision premiums for full-time employees
• 401K plan with a 6% employer match
• Flexibility in schedule and generous paid vacation (available immediately after start date)
• Company-wide holiday office closures
• AD+D and Life Insurance
• Sign-on payments
• Relocation assistance

What This Job Offers

• Job Type: Full-time
• Career Level: Senior