over 1 year ago
Boston, MA, USAMid Level
Responsibilities
- Build computer vision and visual representation learning pipelines for robotic manipulation, including RGB, RGB-D, depth, segmentation, pose, keypoint, and object-centric representations.
- Develop visual models and end-to-end visuomotor policies for reinforcement learning and imitation learning.
- Improve vision-based manipulation data pipelines using domain randomization, photorealistic rendering, synthetic data generation, sensor noise modeling, and real-world fine-tuning.
- Design and train perception models robust to lighting changes, viewpoint shifts, texture variation, clutter, occlusion, object variation, and imperfect calibration.
- Evaluate visual representations and policies on real robotic manipulation tasks, identify failure modes, and iterate on models, data, and training procedures.
- Collaborate with robotics, robot learning, and simulation engineers to define perception strategy for robotic manipulation.
- Set up, calibrate, and evaluate camera and depth-sensing systems and assess their impact on learned policies and real-world robustness.
Requirements
- Ph.D. in computer vision or at least 3 years of experience working on a computer vision product.
- Strong background in machine learning for computer vision, particularly deep learning-based visual perception.
- Experience training modern computer vision models in JAX, PyTorch, or similar frameworks.
- Practical experience with visual representation learning, object detection, segmentation, pose estimation, depth estimation, tracking, or 3D perception.
- Strong Python programming skills.
- Ability to move between research code and production-quality systems.
- Strong understanding of how data distribution, sensor noise, calibration, lighting, and scene variation affect model performance.
- Preferred experience with visual-observation policies, RGB-D data, point clouds, object-centric or learned latent representations, domain randomization, synthetic data generation, differentiable or neural rendering, and photorealistic simulation.
- Preferred experience with Isaac Sim, MuJoCo, or similar robotics simulators and synthetic data tools.
- Preferred familiarity with reinforcement learning, behavior cloning, diffusion policies, offline reinforcement learning, or learning from demonstrations.
- Preferred experience with real-robot deployment, camera and hand-eye calibration, depth sensors, ROS/ROS2, and robot data collection pipelines.
- Preferred first-author publications in venues such as CVPR, ICCV, ECCV, NeurIPS, ICLR, ICML, RSS, CoRL, ICRA, or IROS.
Categories
ML EngineeringRobotics
About Eka
Eka Robotics develops general-purpose robotic systems and autonomy software to bring reliable, high‑speed automation to real‑world environments. The company applies physics‑based methods alongside modern AI to build and deploy robots that can handle varied tasks across industries. Privately held and based in Cambridge, Massachusetts, it focuses on advancing and commercializing robotic intelligence for practical deployment.