Research Engineer/Scientist, Simulation

About the role

You will architect and own the end-to-end simulation pipeline that converts messy real-world facilities into high-fidelity, generative virtual worlds for our wheeled mobile manipulator. You will define what a facility-scale environment looks like in simulation, ensuring that locomotion and manipulation are trained as a single coupled system rather than as isolated behaviors. You will design procedural scene generators that produce challenging layouts and object configurations which expose edge cases in policy behavior across diverse industries. You will build the eval harnesses that measure how well joint base and arm policies perform under photorealistic rendering conditions and noisy perception inputs. You will close the sim-to-real gap by pushing rendering fidelity as a primary first-class output, not a secondary visual polish. You will collaborate tightly with AI Research to identify the most valuable simulated data signals and with Data Ops to decide what must be captured in the real world versus synthesized in simulation. You will operate with founder-level ownership to prioritize simulation investments that directly accelerate the performance of our embodied foundation model across commercial deployments.

Key facts

What you'll do

Design and implement a real-to-sim reconstruction pipeline that ingests raw scans and images of large industrial and retail spaces, producing consistent multi-room 3D models suitable for locomotion and manipulation.

Build photorealistic re-rendering systems that maintain geometric accuracy while elevating visual fidelity to reduce domain discrepancy between synthetic training data and real camera streams.

Create procedural content generation tools that produce navigable facility-scale scenes with variable layouts, dynamic obstacles, and plausible object distributions representative of customer environments.

Develop sim-based data generation loops where failures from real-world locomotion-manipulation episodes are automatically converted into new simulated scenarios for policy retraining.

Construct evaluation suites that stress-test joint base and arm policies under diverse facility configurations, measuring success on locomotion, manipulation, and obstacle avoidance as a unified metric.

Champion rendering as a core simulation deliverable, tuning lighting, materials, and textures to ensure synthetic images closely match real-world visual statistics for vision-based control.

Partner with AI Research to align simulation curriculum with policy learning needs, determining when synthetic data, real-world data, or hybrid datasets best drive generalization improvements.

Collaborate with Data Ops to design storage and streaming architectures that efficiently serve large-scale simulation episodes for training and offline evaluation.

Define and maintain simulation fidelity benchmarks that track how well virtual environments predict real-world policy behavior across hardware iterations.

Lead experiments in domain randomization and asset diversity to ensure policies generalize across unseen facilities without manual scene tuning.

Integrate differentiable rendering techniques where appropriate to enable gradient-based policy optimization running directly in the simulated photorealistic world.

Establish reusable asset libraries for common facility components such as aisles, shelves, doors, and equipment to accelerate scene assembly for new customer deployments.

Drive documentation and internal tooling that lets non-simulation engineers safely iterate on scene parameters and policy configurations without breaking production pipelines.

Champion best practices for reproducibility, versioning, and deterministic simulation runs across research and production workflows.

Requirements

MS or PhD in CS/Robotics/Graphics, or equivalent hands-on experience in related fields.

Demonstrated depth building simulation and rendering systems for robotics, with tangible projects or prototypes to show for it.

Hands-on experience with mainstream simulation stacks such as MuJoCo, Isaac Sim/Isaac Lab, SAPIEN, Omniverse, Blender, or similar platforms.

Proven track record with procedural scene and asset generation, domain randomization, and photorealistic rendering pipelines at real-world scale.

Experience training, evaluating, or deploying manipulation or locomotion policies, and understanding how simulated data feeds into imitation learning, VLA, or diffusion-based training.

Prior work simulating mobile bases, especially wheeled platforms, and appreciation for the differences with tabletop or legged systems.

Background in loco-manipulation or whole-body control that explicitly coordinates mobile-base motion with arm manipulation independent of any particular simulator.

Strong Python programming skills and comfort with PyTorch or JAX when interacting with model training or evaluation workloads.

Senior mindset capable of defining a research agenda and making independent tradeoffs about where simulation effort yields the highest policy performance gains.

Ability to work reliably in a fast-paced startup environment with ambiguous problems and evolving requirements.

Excellent written and verbal communication to coordinate with cross-functional teams including robotics engineers, perception, and operations.

Nice to have

Hands-on experience with real-to-sim-to-real pipelines involving 3D reconstruction, NeRF or Gaussian Splatting, and differentiable rendering.

Exposure to world-model or video-prediction research, such as learned dynamics or latent world models, as a complement to physics-based simulation.

GPU-scale physics simulation experience using CUDA or large-batch parallel simulation rather than single-instance CPU-based sim.

Specific background simulating wheeled mobile manipulators in logistics or warehouse settings, such as Boston Dynamics Stretch or similar platforms.

Publications or public code at venues such as CoRL, RSS, ICRA, NeurIPS, CVPR, or SIGGRAPH that relate to simulation, rendering, or robotic manipulation.