About Mind:

Mind Robotics is building universally capable robots to transform industrial work, starting on the factory floor. We're vertically integrated:

• We train our own foundation models.

• We build our own robots.

• We collect data directly in the factories where our robots will work, starting with Rivian, our first manufacturing partner, in addition to third-party data.

About the team and the role:

Our research speed is set by how fast an idea becomes a training run, and a training run becomes a model on a robot. You'll make that loop fast: large-scale distributed training, data loading from petabyte-scale multimodal datasets, fast evaluation, and getting models running on robots.

Why this is hard:

• The data is physical and big. Synchronized multi-camera video, proprioception, force and tactile streams at petabyte scale. Keeping GPUs fed from affordable storage is a core design problem.

• The model ends on a robot. Inference has real-time latency budgets on compute-limited hardware in a factory.

We're a flat organization, and everyone contributes directly to getting our robots deployed. You'll run experiments alongside the researchers, not just support them.

Responsibilities:

• Scale and harden large-scale distributed training, from parallelism strategy to checkpointing and fault tolerance.

• Profile and speed up the training stack end to end, and track throughput and cost per run.

• Build high-throughput data loading for multimodal robot data, from raw logs to the GPU.

• Make experiments easy to launch, monitor, debug and reproduce, and run your own ablations alongside the researchers.

• Make evaluation fast, and get models running on the robot within its latency budget.

• Build whatever the work needs, and own the problem end to end.

Requirements:

• You've built and scaled training for large models on large multi-node GPU clusters.

• Deep knowledge of PyTorch or JAX and modern distributed training, including what happens under the hood.

• You know where the time goes: you profile, find the bottleneck and fix it.

• Enough understanding of how modern foundation models are trained to shape the experiments, not just run them.

• Strong Python and solid engineering.

Nice to Have:

• Writing custom GPU kernels or working on compilers.

• Petabyte-scale data loading or video decoding on GPU.

• Robotics data or real-time inference on edge hardware.

• Experience as an early hire.

Why Mind

• Your work ships to a factory. Faster training means better robots on a real line, sooner.

• Greenfield. Pick the stack and design it for multimodal robot data from day one.

• Compute isn't the bottleneck. Ample GPU capacity and a petabyte-scale dataset, so your work is about using compute well, not finding it.

• Early ownership. Meaningful equity and a direct say in how we train and deploy models.