Research Engineer, Robotics: Physics Simulation and Sim-to-Real

Persons in these roles are expected to work from our offices in Seattle. On-site requirements vary based on position and team. If you have questions about on-site work arrangements for this role, please ask your recruiter.

Our base salary range is $128,880 - $193,320, and in addition we have generous bonus plans to provide a competitive compensation package.

Who You Are:

You are a simulation engineer who makes physics behave. You have written contact and constraint solvers yourself. When a stack jitters, a grasp slips, an object tunnels through a finger or a gripper linkage explodes, you can tell whether the cause is the timestep, constraint softness, the friction model, the joint formulation or the asset itself, and you fix the root cause rather than adding damping until the problem goes away.

You also know that a stable simulator is not the same as a correct one. You treat logged data from real robots as ground truth. You replay the same motion in simulation, look at the residuals, and close the gap with physics you can explain rather than with magic numbers. You want simulation that a robot foundation model can train on at scale, that never blows up at step ten million, and that behaves like the real world when the policy gets there.

Who We Are:

Ai2’s Robotics team is led by Dieter Fox. We develop scalable simulation domains to study how robot foundation models scale and generalize, and we build robots that are broadly competent in the real world.

Simulation is central to that mission. We train and evaluate policies across MuJoCo, Isaac Sim/Isaac Lab on Newton, and other GPU simulators, and we deploy them on real bimanual and single-arm manipulators. Whether simulated data is worth anything depends on two things: the simulator must be stable enough to run at scale, and close enough to reality that what is learned there transfers. This role owns both.

Our focus areas include manipulation, large-scale training in simulated environments, sim-to-real transfer, robot foundation models and skill learning. We publish at CoRL, ICRA and RSS, as well as CVPR, ICCV, ECCV, NeurIPS, ICLR and ICML, and we release our models, data and tools openly.

We work closely with other teams across Ai2 and with the University of Washington, only 1.5 miles away.

Your Next Challenge:

The needs below are competencies, not a checklist. If you match many of them but not all, we encourage you to apply.

Stability: Make the Simulator Robust at Scale

• Design, implement and improve contact, friction and constraint solvers: complementarity and soft-constraint formulations, pyramidal versus elliptic friction cones, warm-starting, iteration counts and tolerances.

• Find and remove the failure modes that break large-scale training:

• Jitter in resting contact.

• Interpenetration and tunnelling.

• Energy gain.

• Constraint drift in closed kinematic chains.

• Instability under stiff joint drives.

• Non-determinism across runs and GPUs.

• Make articulated robots and closed loops stable and accurate across MuJoCo, Newton and PhysX: parallel grippers, mimic and coupled joints, tendons. Contribute fixes upstream to those open-source engines.

• Choose integrators, timesteps and substeps with evidence. Show how results change with dt, find where a scene stops converging, and quantify the trade-off between accuracy and throughput for GPU-batched training across thousands of environments.

• Build stress and regression tests for physics: long-horizon rollouts, randomized scenes, extreme parameters and contact-rich edge cases, with failures caught before they reach a training run.

• Make assets simulation-ready: collision geometry, inertial properties, joint limits and contact settings that are stable by construction, consistent across URDF, MJCF and OpenUSD.

Fidelity: Bring Simulated Behaviour Closer to Real

• Fit simulation models to logged hardware data: joint friction (Coulomb, viscous, Stribeck), armature and reflected inertia, payload mass and CoM, actuator drive gains, and latency. Identify what the simulator is missing rather than absorbing it into other parameters.

• Calibrate contact behaviour (stiffness, damping, friction, contact dimensionality, compliance) against recorded pushes, slides, grasps and handovers, so objects slip, stick and settle in simulation as they do on the table.

• Keep the same robot consistent across simulators: one source of truth for frames, masses and contact settings, with MuJoCo, Newton and PhysX agreeing to within millimetres, and every remaining gap explained.

• Own a sim-to-real correlation suite that runs like a test suite:

• Replay recorded trajectories in every simulator.

• Report per-joint RMS, onset lag, velocity profiles and contact-force error against thresholds set in advance.

• Catch regressions whenever a solver, asset or parameter changes.

• Turn measured uncertainty into domain-randomisation ranges for policy training, and measure whether a policy’s real-world success tracks its success in the calibrated simulator.

• Write a model card for each robot and scene: every parameter, where it came from and how much to trust it.

What You’ll Need:

Minimum Qualifications:

• PhD in robotics, mechanical engineering, applied mathematics, physics, computer graphics or a related field, or a Masters + 3 years of experience. Exceptional candidates with a bachelor’s degree and 3+ years of relevant industry experience are also welcome to apply.

• Demonstrated depth in rigid-body dynamics and constrained multibody systems: articulations (reduced and maximal coordinates), contact and friction modelling, and numerical integration of stiff systems.

• Solver or physics-engine code you wrote yourself, in a research codebase, a production engine, or a widely used open-source simulator such as MuJoCo, Newton/Warp, PhysX, Drake, Bullet or Genesis.

• A track record of diagnosing and fixing simulation instability (jitter, penetration, energy drift, divergence), with an understanding of why the fix works.

• Experience fitting simulation parameters to measured data and validating the result quantitatively against real-world recordings.

• Proficiency in Python and C++, experience with robot description formats (URDF, MJCF, OpenUSD), and knowledge of at least one major simulation stack at source level.

Preferred Qualifications:

• GPU programming for physics (CUDA, NVIDIA Warp or JAX) and experience scaling simulation for reinforcement or imitation learning.

• Experience with deformable, cable or FEM simulation (XPBD, VBD, corotational or hyperelastic FEM, Cosserat rods) and its coupling with rigid bodies.

• Familiarity with differentiable simulation, Bayesian or gradient-based parameter estimation, or learned residual dynamics.

• Experience in robot manipulation, grasping or contact-rich tasks, in simulation or on hardware.

• Publications or open-source contributions in simulation, dynamics or sim-to-real transfer (CoRL, RSS, ICRA, SIGGRAPH/SCA, or engine repositories).

Physical Demands and Work Environment:

The physical demands described here are representative of those that must be met by a team member to successfully perform the essential functions of this position. Reasonable accommodations may be made to enable individuals with disabilities to perform the functions.

• Must be able to remain in a stationary position for long periods of time.

• The ability to communicate information and ideas so others will understand. Must be able to exchange accurate information in these situations.

• The ability to observe details at close range.

• Can work under deadlines.

A Little More About Ai2:

Ai2 is a Seattle based non-profit AI research institute founded in 2014 by the late Paul Allen. Our mission is building breakthrough AI to solve the world’s biggest problems. We develop foundational AI research and innovation to deliver real-world impact through large-scale open models, data, robotics, conservation, and beyond.

…