Job description

Dyna Robotics builds general-purpose robots powered by a proprietary embodied AI foundation model with top-in-industry generalization and real-world performance. Already deployed with customers across multiple industries, our robots do commercial-grade work in the physical world. Our team comes from Google DeepMind, Meta, and Cruise, and we're backed by CRV, First Round, and other leading investors. THE ROLE Most ML infrastructure work ends at a benchmark. Ours ends at a robot doing real work at a customer site. You'll own our training and inference infrastructure end to end, from the moment robot data lands to the moment a new model runs on hardware. The goal: every GPU busy, every run reproducible, and every researcher's next experiment one command away. WHY THIS IS HARDER THAN TYPICAL ML INFRA - The data is physical: synchronized video, proprioception, and 3D signals from real robots, not clean text. - Latency is physical too: a slow model means a robot that hesitates or misses a grasp. - The loop is closed: data from deployed robots feeds back into training, and your systems sit at the center of that flywheel. WHAT YOU’LL DO - Scale distributed training across a multi-cloud GPU fleet using sharding, activation checkpointing, and memory optimization (FSDP, ZeRO). - Build a research codebase and scheduling system (Kubernetes/SLURM) built for fast iteration, automatic retries, and painless failure recovery. - Design high-throughput pipelines for terabytes of multimodal robot data so dataloaders never starve the GPUs. - Ship low-latency inference for real-time robot control using quantization, distillation, and compilation (TensorRT, Triton). - Profile GPU utilization, I/O bottlenecks, and memory fragmentation to get the most out of every node. YOU'LL THRIVE HERE IF YOU - Genuinely care about embodied AI. You follow robot learning research, have opinions about VLAs or world models, or have built something that moves. You want your work to show up in the physical world, not just on a dashboard. - Have 7+ years of engineering experience, including leading projects in HPC or ML infrastructure. - Know PyTorch and distributed training deeply (DeepSpeed, Accelerate, FSDP), including mixed precision and gradient accumulation. - Have hands-on experience with cloud GPU environments (GCP/AWS) and Kubernetes. - Understand distributed systems at a low level: race conditions, memory management, NCCL and inter-node communication. - Own what you ship: you design, build, and operate systems end to end to unblock fast-moving research. BONUS POINTS FOR - Robotics data formats (MCAP, Protobuf) or multimodal models such as VLAs. - Custom kernels (Triton), compilers, or runtime optimization. - Experience as a founding or early infrastructure hire. - Side projects involving robots, simulators, or hardware. At Dyna Robotics, we build technology for the real world, which requires a team as diverse as the environments our robots inhabit. We are an equal opportunity employer committed to technical rigor and mutual respect. Don’t let a checklist stop you. Data shows that underrepresented groups often only apply if they meet 100% of the criteria. We value problem-solving and grit over keyword matching. If you’re passionate about robotics, no matter your discipline, we want to hear from you, even if you don't check every box.

Requirements

Must-haves

Most ML infrastructure work ends at a benchmark. Ours ends at a robot doing real work at a customer site.
WHY THIS IS HARDER THAN TYPICAL ML INFRA
The data is physical: synchronized video, proprioception, and 3D signals from real robots, not clean text.
Latency is physical too: a slow model means a robot that hesitates or misses a grasp.
The loop is closed: data from deployed robots feeds back into training, and your systems sit at the center of that flywheel.
Scale distributed training across a multi-cloud GPU fleet using sharding, activation checkpointing, and memory optimization (FSDP, ZeRO).
Build a research codebase and scheduling system (Kubernetes/SLURM) built for fast iteration, automatic retries, and painless failure recovery.
Design high-throughput pipelines for terabytes of multimodal robot data so dataloaders never starve the GPUs.
Ship low-latency inference for real-time robot control using quantization, distillation, and compilation (TensorRT, Triton).
Profile GPU utilization, I/O bottlenecks, and memory fragmentation to get the most out of every node.
Have 7+ years of engineering experience, including leading projects in HPC or ML infrastructure.
Know PyTorch and distributed training deeply (DeepSpeed, Accelerate, FSDP), including mixed precision and gradient accumulation.
Have hands-on experience with cloud GPU environments (GCP/AWS) and Kubernetes.
Understand distributed systems at a low level: race conditions, memory management, NCCL and inter-node communication.
Own what you ship: you design, build, and operate systems end to end to unblock fast-moving research.
Robotics data formats (MCAP, Protobuf) or multimodal models such as VLAs.
Custom kernels (Triton), compilers, or runtime optimization.
Experience as a founding or early infrastructure hire.
Side projects involving robots, simulators, or hardware.
At Dyna Robotics, we build technology for the real world, which requires a team as diverse as the environments our robots inhabit. We are an equal opportunity employer committed to technical rigor and mutual respect.

Nice-to-haves

None flagged.

Application prep

Use the role’s language and application details while tailoring your materials.

Keywords to mirror

aimlkubernetesgcpaws

Platform: ashby
Effort: quick