Preparing for a Research Scientist interview at & General Intuition requires a balanced approach. You must demonstrate deep theoretical mastery of machine learning, robotics, or computer vision while proving that you are a highly capable engineer who can write clean, scalable production code.

Research Depth & Core ML Foundations – You must be able to explain the "why" behind your technical choices. Interviewers will push you on the mathematical formulations of your past work, the limitations of your approaches, and alternative methods you considered. Be prepared to defend your research decisions from first principles.

Systems & Engineering Rigor – Excellent research is useless if it cannot be implemented efficiently. You will be evaluated on your ability to write high-quality, vectorized, and performant Python/PyTorch code. You should be comfortable discussing computational complexity, GPU memory management, and distributed training paradigms.

Problem-Solving & First-Principles Thinking – When faced with ambiguous, open-ended physical AI problems, your ability to structure your thoughts is critical. Break down complex systems into manageable components, state your assumptions clearly, and walk your interviewer through your iterative design process.

Collaborative Execution & Communication – Building physical AI systems is a team sport. You must demonstrate that you can collaborate effectively with software engineers, hardware teams, and product managers. Clearly articulating complex technical concepts to non-specialists is highly valued.

The interview process for a Research Scientist at & General Intuition is designed to evaluate both your scientific creativity and your engineering execution. It is structured to ensure that you possess the depth required to push the boundaries of AI, as well as the practical skills to deploy models onto physical hardware.

The process typically begins with a recruiter screen to discuss your background, research interests, and alignment with the company's mission. This is followed by one or two technical phone screens, focusing on core ML theory, live coding (usually in PyTorch), and a deep dive into your past publications or projects. If you pass these initial stages, you will proceed to a comprehensive onsite interview panel consisting of research presentations, system design sessions, and behavioral evaluations.

The onsite panel is highly collaborative and rigorous. You will interact directly with senior researchers and engineering leads who will challenge your assumptions and explore how you think under pressure. The team values candidates who are intellectually honest, receptive to feedback during the interview, and eager to solve highly complex, ambiguous problems.

This timeline outlines the typical path from your initial application to the final offer. The process generally spans 3 to 5 weeks, moving from high-level alignment to intense technical evaluation, culminating in a comprehensive onsite panel that tests both your research and engineering capabilities.

To excel in the & General Intuition interview process, you must perform exceptionally well across several core evaluation areas. Below is a detailed breakdown of what to expect and how to prepare for each.

Reinforcement Learning & Control

This evaluation area focuses on your ability to design decision-making agents that can operate safely and efficiently in dynamic physical environments.

Be ready to go over:

• Policy Optimization – Deep understanding of TRPO, PPO, SAC, and DDPG, including their mathematical derivations, objective functions, and practical hyperparameters.
• Reward Engineering – How to structure sparse vs. dense rewards, handle multi-objective optimization, and implement reward shaping safely.
• Offline RL & Imitation Learning – Algorithms like CQL, IQL, and Behavior Cloning, and how to leverage offline datasets to bootstrap policy training.
• Advanced concepts (less common) – Hierarchical reinforcement learning, meta-RL, and multi-agent decision-making under partial observability.

Example questions or scenarios:

• "How would you design an RL agent to perform aggressive lane-merging maneuvers in heavy highway traffic, ensuring both safety and passengers' comfort?"
• "Explain how you would debug a policy that performs exceptionally well in simulation but consistently fails when deployed on the physical robot."

3D Vision, Perception & Generative World Models

Here, you will be evaluated on your ability to reconstruct, understand, and simulate the 3D physical world using multi-modal sensor inputs.

Be ready to go over:

• 3D Representations – The trade-offs between implicit neural representations (NeRFs, SDFs) and explicit representations (point clouds, meshes, 3D Gaussians).
• Generative Modeling – Diffusion models, GANs, and autoregressive models applied to video generation, sensor simulation, and closed-loop world modeling.
• Sensor Fusion – Combining camera, LiDAR, and radar data using spatial-temporal transformers or bird's-eye-view (BEV) networks.
• Advanced concepts (less common) – Differentiable rendering, self-supervised 3D pre-training, and open-vocabulary 3D scene understanding.

Example questions or scenarios:

• "Design a generative world model that can predict the next 5 seconds of a driving scene given the current camera inputs and the vehicle's planned actions."
• "How would you build a real-time 3D reconstruction pipeline for a robotic manipulator using only two low-cost RGB cameras?"

Sim-to-Real & Robotics Systems

This area tests your practical understanding of deploying machine learning models onto physical hardware, bridging the gap between theory and reality.

Be ready to go over:

• Domain Randomization & Adaptation – Systematic techniques to randomize visual assets, physics parameters, and dynamics to ensure robust transfer.
• System Identification – Methods to align simulation parameters with physical hardware measurements.
• Hardware Constraints – Designing models that respect strict latency budgets, memory bandwidth limits, and safety boundaries.
• Advanced concepts (less common) – Online adaptation, residual policy learning (combining ML with classical control), and tactile sensor integration.

Example questions or scenarios:

• "A policy trained in simulation to grab objects fails on the physical arm due to unexpected friction and joint backlash. How do you diagnose and resolve this without retraining from scratch?"
• "Describe how you would design a safe fallback controller that overrides an RL policy when it detects a high-probability collision state."