New opportunity

Machine Learning / Reinforcement Learning Engineer

Eka Robotics · Cambridge, United States

About this role

Eka Robotics is hiring an on-site Machine Learning / Reinforcement Learning Engineer in the Boston Area to develop and evaluate learning-based policies for robotic manipulation. The structured job location identifies Cambridge, Massachusetts, United States.

Responsibilities

Research and implement reinforcement learning and supervised learning algorithms for robotic manipulation.
Design simulation models and domain randomization strategies with the robotics team to align simulations with physical systems.
Run experiments to optimize model architectures, sample complexity and policy performance while meeting real-time execution constraints.
Build scalable pipelines for real and synthetic data, and select algorithms to improve data efficiency and policy performance.
Deploy, evaluate and debug policies on physical robots; identify bottlenecks and implement improvements with the robotics team.

Qualifications

BS, MS or PhD in Computer Science, Robotics or a related field; deep theoretical and practical knowledge of reinforcement learning and supervised learning.
Experience with physics engines such as Isaac Sim, MuJoCo or PyBullet, and robotics middleware such as ROS or ROS2.
Understanding of Transformers, CNNs and foundation models; expert-level Python and proficiency with a deep learning framework such as PyTorch or JAX.
Commitment to clean code, version control and reproducible experiments. A record of publications at top-tier robotics or machine learning conferences, or a project portfolio demonstrating strong practical expertise, is required.

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

Skills for this role

Reinforcement LearningSupervised LearningRobotic ManipulationSimulationDomain RandomizationModel OptimizationData PipelinesPythonIsaac SimMuJoCoPyBulletROSROS2TransformersCNNsFoundation ModelsPyTorchJAXVersion ControlExperimental Design

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

Develop manipulation policies
The role calls for researching and implementing reinforcement and supervised learning algorithms for robotic manipulation. Prepare to explain how you choose and evaluate an approach for a manipulation task.
Build simulations that support transfer
The posting specifies simulation models and domain randomization developed with the robotics team to align simulation with physical systems. Prepare to discuss how you would test whether that alignment is useful.
Optimize policies through experiments
The role requires experiments on model architecture, sample complexity, and policy performance under real-time execution constraints. Prepare to explain your metrics, baselines, and reproducible experiment design.
Engineer efficient data workflows
The posting calls for scalable pipelines for real and synthetic data and algorithm choices that improve data efficiency. Prepare to describe how data quality, provenance, and pipeline reliability affect comparisons.
Close the loop on physical robots
The engineer will deploy, evaluate, and debug policies on hardware, then address bottlenecks with the robotics team. Prepare to explain how you isolate a policy issue from an integration or evaluation issue.

Likely challenges to think through

Simulation fidelity versus useful variation
A likely challenge, not a confirmed employer problem, is that domain randomization may improve robustness while obscuring which simulation mismatch matters. Prepare an example of measuring transfer on hardware and refining simulation from observed failures.
Policy quality versus execution time
A likely trade-off is that a more capable architecture may require more data or run too slowly for real-time use. Prepare to compare policy quality, sample complexity, and execution time with explicit measurement criteria.
Research iteration versus reliable hardware evaluation
A likely difficulty is separating learning gains from changes in data, software, or robot behavior during physical tests. Prepare a reproducible evaluation and debugging approach with versioned experiments and clear bottleneck measurements.

Practice your answers

Technical practice

How would you choose between reinforcement learning and supervised learning for a robotic manipulation policy?

What to coverSuggested answer focus: Compare the data and feedback each approach needs, and explain when demonstrations or task rewards would be more useful. Describe how you would evaluate the resulting policy on manipulation performance and data efficiency.

How would you improve the sample efficiency of a reinforcement learning approach?

What to coverSuggested answer focus: Explain a plausible way to make better use of collected experience, then identify the assumptions it introduces. Describe an experiment that compares policy performance against the amount of data used.

How would you design a simulation model for a manipulation task?

What to coverSuggested answer focus: Identify physical interactions that matter to the policy and explain how you would check whether the simulation reflects observations from the robot. Discuss the trade-off between model fidelity and iteration speed in a physics engine such as Isaac Sim, MuJoCo, or PyBullet.

How would you design and evaluate domain randomization?

What to coverSuggested answer focus: Explain which simulation parameters you would consider varying and how you would choose ranges grounded in the physical system. Compare simulated and on-robot results to detect randomization that is too narrow or that makes training unnecessarily difficult.

How would you compare CNNs, Transformers, and foundation models for a manipulation policy?

What to coverSuggested answer focus: Relate each architecture to the available inputs, training data, and desired policy behavior without assuming one is always best. Explain how you would test performance and execution cost under real-time constraints.

How would you design an experiment to compare policy architectures?

What to coverSuggested answer focus: Define measures for policy performance, sample complexity, and execution time, and keep comparison conditions consistent. Explain how you would interpret a performance gain that comes with higher data use or slower inference.

How would you build a scalable pipeline for real and synthetic robot data?

What to coverSuggested answer focus: Describe how you would organize, validate, and track data so experiments can distinguish its source and reproduce training inputs. Discuss the trade-off between collecting more data and improving the usefulness of existing data.

How would you diagnose a policy that works in simulation but fails on a physical robot?

What to coverSuggested answer focus: Separate possible simulation mismatch, input or output integration issues, and execution-time bottlenecks before changing the model. Explain how controlled on-robot evaluations and collaboration with the robotics team would help narrow the cause.

What would you check when integrating a learned policy through ROS or ROS2?

What to coverSuggested answer focus: Discuss how you would verify that observations and actions are communicated as intended and that the policy meets its execution-time constraint. Explain how you would distinguish a middleware or timing issue from a policy-quality issue.

How would you make Python-based policy experiments reproducible?

What to coverSuggested answer focus: Explain how clean code, version control, and recorded experiment settings support reliable comparisons in PyTorch or JAX. Describe how another researcher could trace a reported result back to its code, data, and evaluation.

Behavioural practice

Tell me about a time you aligned a simulation with a physical system alongside a robotics team.

What to coverSuggested answer focus: Give the situation and the disagreement or mismatch you needed to resolve, then explain your actions with the team. State the observed result and what evidence showed whether alignment improved.

Tell me about a time an experiment challenged your preferred approach.

What to coverSuggested answer focus: Describe the situation, your initial hypothesis, and the evidence that contradicted it. Explain the action you took next and the resulting decision or learning.

Tell me about a time you debugged a learned system on physical hardware.

What to coverSuggested answer focus: Set out the observed failure, how you isolated its cause, and any collaboration needed to make a change. Report the result using evidence from evaluation rather than assuming the change worked.

Tell me about a time you improved the reproducibility of a research or engineering workflow.

What to coverSuggested answer focus: Describe the original workflow problem and the specific changes you made to code, versioning, or experiment records. Explain the result for later comparisons or for other contributors.

Tell me about a research result or project that demonstrated practical robotics or machine learning expertise.

What to coverSuggested answer focus: Give the situation and your own contribution, including the technical choice you made and why. Describe the outcome, its limitations, and the evidence supporting your conclusion.

Sharpen your knowledge

Topics to prepare

Reinforcement and supervised learning for manipulation
Posting: Deep theoretical and practical knowledge of both is required for robotic manipulation. Suggested exercise: compare when to train a policy from labeled demonstrations versus interaction, and explain how each choice affects data needs and policy evaluation.
Simulation fidelity and domain randomization
Posting: The role designs simulation models and randomization strategies aligned with physical systems. Suggested exercise: list physical parameters that could vary in a manipulation simulation, then explain how you would test whether the resulting policy transfers to hardware.
Architecture trade-offs
Posting: Deep understanding of Transformers, CNNs, and foundation models is requested. Suggested exercise: compare their possible roles in a manipulation policy and explain how you would assess performance against real-time execution constraints; the comparison is preparation, not a prescribed design.
Sample complexity and policy evaluation
Posting: Experiments must assess model architectures, sample complexity, policy performance, and real-time execution. Suggested exercise: sketch a comparison with a fixed training budget, clearly defined performance measures, and an execution-time measurement.
Data efficiency across real and synthetic data
Posting: The role manages both data sources and selects algorithms to improve data efficiency. Suggested exercise: explain how you would compare learning curves from real and synthetic data while checking whether gains persist during physical evaluation.

Skills to revise

Implement a learning baseline
Posting: Expert-level Python and proficiency with PyTorch or JAX support RL and supervised-learning implementation. Suggested exercise: implement a small manipulation-policy training baseline in Python with one named framework, then inspect training behavior and evaluation results.
Build and vary a physics simulation
Posting: Experience with physics engines such as Isaac Sim, MuJoCo, or PyBullet and domain randomization is requested. Suggested exercise: in one listed engine, vary a manipulation-relevant physical parameter and compare policy results across settings.
Prepare real and synthetic data
Posting: The role develops scalable data-management pipelines for both sources. Suggested exercise: design a Python pipeline that labels data provenance, checks input consistency, and produces repeatable training and evaluation splits.
Integrate and debug a policy with robotics middleware
Posting: ROS or ROS2 experience and on-robot policy deployment, evaluation, and debugging are specified. Suggested exercise: trace policy inputs and outputs through a ROS or ROS2 integration, then prepare a bottleneck checklist for physical evaluation; a simulated walkthrough is an option if hardware is unavailable.
Run reproducible experiments
Posting: Clean code, version control, and reproducible workflows are qualifications. Suggested exercise: version a training change, record its configuration and results, and rerun a baseline to verify the comparison before drawing conclusions.

Your preparation checklist

Prepare a manipulation-project walkthrough
Choose a publication or portfolio project, as the posting requests evidence of practical expertise. Prepare a concise account of the task, your contribution, methods, results, limitations, and what you would change.
Compare learning approaches
The role covers reinforcement and supervised learning. Prepare a concrete example explaining why you chose one approach over another, including its data needs and failure modes.
Sketch a simulation-to-robot design
As a preparation exercise, outline a manipulation simulation, domain-randomization choices, and how you would check alignment with a physical system. Be ready to discuss relevant physics-engine and ROS/ROS2 experience.
Draft a policy experiment plan
Specify baselines, metrics, and controlled comparisons for architecture choice, sample complexity, policy performance, and real-time execution. Prepare to explain where CNNs, Transformers, or foundation models would—or would not—help.
Show a reproducible data workflow
Prepare a small diagram or code example covering real and synthetic data, Python with PyTorch or JAX, version control, and the records needed to reproduce an experiment.
Rehearse an on-robot debugging case
Walk through how you would deploy and evaluate a policy, isolate a performance bottleneck, and work with robotics teammates on an improvement. Distinguish any real experience from a hypothetical approach.
Write targeted questions for Eka Robotics
Ask how the team measures manipulation-policy success, checks simulation-to-hardware alignment, manages real versus synthetic data, and defines real-time constraints; the posting names these responsibilities but does not give their implementation details.

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…