Data Science

The Role

We're hiring a Data Scientist with deep experience in video analytics, action understanding, robot-learning data preprocessing, and multimodal pipelines. This role sits at the core of our product: extracting signal from chaotic, real-world sensor streams. You will be shaping how humanoids and manipulators learn from human demonstrations across factories, warehouses, hospitality sites, and more.

What You'll Do

• •Build and optimize video + sensor processing pipelines for egocentric and teleop data.
• •Develop segmentation, action-chunking, temporal alignment, and quality-filtering algorithms for robotic imitation learning.
• •Analyze multimodal datasets (RGB, depth, IMU, tactile, robot states) and identify failure modes, drift, and noise.
• •Work with an ML team to prepare ACT/FAST-tokenized or similar representations for policy training.
• •Create internal tools for dataset health monitoring, automated scoring, and evaluator model validation.
• •Stress-test assumptions: what signals matter? what can be thrown away? what's actually learnable?

Who You Are

• •You think in systems, not notebooks.
• •Strong in Python, PyTorch, CV/ML, and familiar with robotics-adjacent tooling (ROS, teleop systems, sensor sync).
• •Comfortable with bad data, missing timestamps, corrupt frames, and noisy sensors — and extracting order from them.
• •Prefer measurable impact over academic elegance; skeptical of magic-model thinking.
• •Bonus: experience with imitation learning (ALOHA, ACT-style models), action segmentation, or multimodal transformer pipelines.

Why Humaid

• •Work directly on the bottleneck of Physical Intelligence — real-world data.
• •Access to large industrial environments (+35,000 workers) where we run live pilots and capture rare edge-cases at scale.
• •Fast, practical team that actually deploys — not a research museum.
• •Our scientific direction is led by Chief Science Officer Ruslan Salakhutdinov (ex-VP of AI Research @ Meta; Professor at Carnegie Mellon)