New opportunity

Machine Learning / Computer Vision Engineer

Eka Robotics · Cambridge, United States

About this role

Eka Robotics seeks a Machine Learning / Computer Vision Engineer to develop visual perception for robotic manipulation. The posting lists the location as Boston Area; its structured job location specifies Cambridge, Massachusetts, United States. The role is full-time and on-site.

Responsibilities

Build computer vision and visual representation learning pipelines using RGB, RGB-D, depth, segmentation, pose, keypoint and object-centric representations. Develop visual models for reinforcement learning and imitation learning, including end-to-end policies that map visual observations to robot actions.
Improve data pipelines through domain randomization, photorealistic rendering, synthetic data generation, sensor noise modeling and real-world fine-tuning. Train perception models to withstand changes in lighting, viewpoint, texture and object instances, as well as clutter, occlusion and imperfect calibration.
Evaluate representations and policies on real robotic manipulation tasks, diagnose failures and refine models, data and training. Work with robotics, robot learning and simulation engineers on perception strategy; set up, calibrate and assess camera and depth-sensing systems when needed.

Minimum qualifications:

A Ph.D. in computer vision or 3+ years of experience working on a computer vision product. Strong machine learning and deep learning-based visual perception skills, experience training computer vision models with JAX, PyTorch or similar frameworks, and strong Python programming skills.
Practical experience in at least one of visual representation learning, object detection, segmentation, pose estimation, depth estimation, tracking or 3D perception. Ability to work across research code and production-quality systems, with an understanding of how data distribution, sensor noise, calibration, lighting and scene variation affect performance.

Preferred qualifications

Experience training policies from visual observations; using domain randomization, synthetic data, differentiable or neural rendering, or photorealistic simulation; or working with Isaac Sim, MuJoCo or similar tools.
Familiarity with reinforcement learning, behavior cloning, diffusion policies, offline RL or learning from demonstrations. Real-robot deployment experience involving calibration, depth sensors, ROS/ROS2 or robot data collection is valued, as are first-author publications at leading computer vision, robotics or machine learning venues.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Computer VisionMachine LearningDeep LearningVisual Representation LearningRobotic ManipulationRGB-DObject DetectionSegmentationPose EstimationDepth EstimationObject Tracking3D PerceptionReinforcement LearningImitation LearningVisuomotor PoliciesDomain RandomizationSynthetic Data GenerationPhotorealistic RenderingSensor Noise ModelingCamera CalibrationDepth SensingJAXPyTorchPythonIsaac SimMuJoCoROSROS2Behavior CloningDiffusion Policies

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Build perception pipelines for manipulation
The role calls for visual representation learning pipelines using inputs and representations such as RGB-D, depth, segmentation, pose, keypoints, and object-centric features. Prepare to explain why a representation is useful for a particular manipulation task, not just its vision accuracy.
Connect vision to robot actions
Developing visual models for reinforcement learning and imitation learning, including end-to-end visuomotor policies, is an explicit responsibility. Prepare to trace how observations become policy inputs and how visual errors could affect actions.
Improve robustness through data and training
The posting emphasizes domain randomization, rendering, synthetic data, sensor-noise modeling, real-world fine-tuning, and resilience to lighting, viewpoint, clutter, occlusion, and imperfect calibration. Prepare to explain how you would isolate which variation is causing a failure.
Iterate on real-robot results with partner teams
The engineer is expected to evaluate representations and policies on real manipulation tasks, diagnose failures, and collaborate with robotics, robot learning, and simulation engineers. Camera and depth-sensor setup and evaluation are also part of the role when needed.

Likely challenges to think through

Synthetic coverage versus real-world fidelity
A likely trade-off is whether broader randomized data captures useful variation or misses sensor and scene effects that matter on a robot; this is a preparation hypothesis, not a reported problem at Eka. Prepare a way to compare synthetic and real failure cases and decide when to adjust generation versus fine-tune on real data.
Vision metrics versus manipulation outcomes
A representation can perform well on a perception benchmark yet supply inadequate information for a policy. As a preparation hypothesis, work through how you would compare representation-level results with real-task outcomes and diagnose whether failures arise in perception, policy learning, or their interface.
Sensor choices versus robustness
Choosing and calibrating RGB or depth sensing may improve one operating condition while increasing sensitivity to viewpoint, noise, or imperfect calibration. This is an inferred trade-off. Prepare to explain how you would evaluate sensor configurations under varied scenes and communicate the results to robotics and learning collaborators.

Practice your answers

Technical practice

How would you choose between RGB, RGB-D, point clouds, and object-centric representations for a manipulation task?

What to coverThe role involves visual representations for robotic manipulation, including these observation types. Suggested preparation: compare what each captures, its sensing and computation costs, and how you would test whether it helps the robot act reliably.

When would you use segmentation, pose, or keypoints rather than a learned visual representation?

What to coverThe posting names these perception outputs alongside visual representation learning. Suggested preparation: discuss the information each exposes to a policy, where errors or occlusion could matter, and how you would compare downstream task performance.

How would you train and evaluate a policy that maps visual observations to robot actions?

What to coverDeveloping end-to-end visuomotor policies is a stated responsibility. Suggested preparation: explain the observation-to-action training setup, then distinguish perception metrics from success and failure on real manipulation tasks.

How would you choose among behavior cloning, reinforcement learning, and other demonstration-based methods for a visual policy?

What to coverThe role supports reinforcement and imitation learning; behavior cloning, offline RL, and diffusion policies are preferred areas of familiarity. Suggested preparation: compare their data needs, feedback or supervision, and likely failure modes without assuming the team uses one particular method.

How would you use synthetic data and domain randomization to improve real-world performance?

What to coverThe data-pipeline responsibilities include synthetic generation, photorealistic rendering, and domain randomization. Suggested preparation: describe what scene and appearance factors you would vary and how real-task evaluation would reveal whether the changes transfer.

How would you model sensor noise and decide what real-world fine-tuning data to collect?

What to coverSensor noise modeling and real-world fine-tuning are explicitly part of the role. Suggested preparation: explain how you would compare synthetic and real observations, prioritize mismatches that affect policy behavior, and check whether fine-tuning helps beyond the collected examples.

How would you diagnose a perception model that fails under occlusion, lighting changes, or new object instances?

What to coverThe posting calls for robustness to these conditions, as well as clutter, viewpoint shifts, and imperfect calibration. Suggested preparation: propose condition-specific tests that separate data-distribution problems from model limitations, then explain how results would guide retraining.

How would camera, depth-sensor, and hand-eye calibration affect a learned manipulation policy?

What to coverThe role includes evaluating camera and depth systems when needed, and hand-eye calibration is listed as preferred experience. Suggested preparation: explain how sensing and calibration errors propagate into visual observations or actions, and how you would detect and assess their effect on real tasks.

How would you evaluate whether a new visual representation improves real-robot manipulation?

What to coverEvaluating representations and policies on real tasks and iterating on failures are stated responsibilities. Suggested preparation: describe a comparison that holds the task and evaluation conditions consistent, examines failure categories, and checks whether gains persist across scene variation.

How would you take a vision-training experiment toward a production-quality system?

What to coverThe minimum qualifications call for model training with JAX, PyTorch, or similar frameworks, strong Python skills, and movement between research code and production-quality systems. Suggested preparation: discuss reproducibility, data and model checks, and trade-offs between experimental speed and dependable operation.

Behavioural practice

Tell me about a time you aligned perception work with robotics, robot-learning, or simulation collaborators.

What to coverThe role calls for collaboration on perception strategy. Suggested preparation: use a real situation, your actions to resolve differing technical priorities, and the result for the shared work.

Tell me about a time real-world testing revealed a failure your earlier evaluation missed.

What to coverThe posting emphasizes diagnosing failures on real manipulation tasks and iterating on models, data, and training. Suggested preparation: explain the situation, how you investigated and changed course, and what the subsequent results showed.

Tell me about a time you moved an experimental vision model toward reliable use.

What to coverThe role requires working across research code and production-quality systems. Suggested preparation: describe the initial situation, the actions you took to improve reliability, and the observed result, including any remaining trade-offs.

Tell me about a time you had to choose what data or sensing issue to address first.

What to coverThe posting connects data distribution, sensor noise, calibration, and scene variation to model performance. Suggested preparation: give a real situation, the evidence behind your prioritization and actions, and the result; do not assume a particular sensor setup for this role.

Sharpen your knowledge

Topics to prepare

Visual representations for manipulation
Compare when RGB, RGB-D, depth, segmentation, pose, keypoints, and object-centric representations preserve information needed for grasping under occlusion. The role explicitly builds these perception pipelines.
Robustness under scene and sensor variation
For lighting shifts, clutter, viewpoint changes, sensor noise, and imperfect calibration, predict distinct failure modes and suitable evaluation slices. The posting explicitly requires understanding how these factors affect model performance.
Visual observations to robot actions
Sketch how an image-based representation feeds an imitation-learning or reinforcement-learning policy, then identify what an end-to-end policy might gain or lose. Developing visuomotor models is a stated responsibility; specific robot-learning methods are preferred familiarity.
Synthetic-to-real data strategy
Explain which variations domain randomization and sensor-noise modeling can cover, and which would still call for real-world fine-tuning. The role explicitly includes synthetic data, rendering, and real-world robustness.
Camera geometry and depth sensing
Trace how a camera or hand-eye calibration error could propagate from a depth observation into an incorrect robot action. Evaluating camera and depth systems is a stated responsibility when needed; hand-eye calibration experience is preferred.

Skills to revise

Train and diagnose a vision model
In Python with JAX, PyTorch, or a similar framework, rehearse a small detection, segmentation, or pose-training workflow and inspect errors by object instance and occlusion. Model training, Python, and practical experience in at least one listed perception area are minimum qualifications.
Build a multimodal perception pipeline
As a practice exercise, outline code that aligns RGB and depth inputs, produces a keypoint or object-centric representation, and checks missing or noisy depth. The role explicitly calls for RGB-D and visual-representation pipelines.
Run a robustness evaluation
Draft a test matrix varying lighting, viewpoint, clutter, and calibration; record both perception errors and downstream manipulation failures, then propose one data or model change. The posting calls for evaluating real-task failures and iterating on models, data, and training.
Harden a research-to-production workflow
Practice turning a training or inference prototype into a reproducible Python pipeline with input validation and clear failure reporting. The minimum qualifications explicitly require moving between research code and production-quality systems.
Exercise a visual-policy data workflow
As preparation for a preferred area, outline how to collect visual observations and robot actions, train a behavior-cloning baseline, and compare outcomes across scene variations. The posting values training policies from visual observations but does not require behavior cloning specifically.

Your preparation checklist

Map your evidence to the qualifications
Prepare a one-page evidence sheet for the stated Ph.D.-or-3+-years computer-vision-product qualification, Python and model training in JAX, PyTorch or a similar framework. Identify concrete work in at least one listed perception area; distinguish minimum qualifications from preferred experience.
Build a perception project walkthrough
Choose a relevant project and prepare a short walkthrough of its visual inputs, representation or prediction task, training data, metrics, your contributions and failure cases. The role specifically covers perception pipelines for robotic manipulation.
Practice a visuomotor design exercise
Sketch how RGB, RGB-D or object-centric observations could feed a manipulation policy, and explain the trade-offs between a separate perception model and an end-to-end policy. Treat this as interview practice, not a stated hiring exercise.
Prepare a robustness and sim-to-real test plan
Outline tests for lighting, viewpoint, clutter, occlusion, object variation and imperfect calibration. Explain how you would use synthetic data, domain randomization or sensor-noise modeling, then check whether real-world fine-tuning improves task performance.
Document an iteration from failure to fix
Prepare one example showing how you diagnosed a perception or policy failure, changed data, models or training, and measured the result. If available, include real-manipulation results; the posting emphasizes evaluation on real robotic tasks.
Explain sensor and engineering trade-offs
Prepare a concise account of camera or depth-sensor selection and calibration, including hand-eye calibration if relevant, and how sensor errors might affect a learned policy. Bring an example of moving Python research code toward a production-quality system.
Write targeted questions for Eka Robotics
Ask which manipulation tasks and sensors the role prioritizes, how learned policies are evaluated on real robots, and how perception work is divided among robotics, robot-learning and simulation engineers. These are questions to clarify, not assumed company processes.

Availability last checked: Nov 4, 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in United States. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…