Role description
Destro AILocation: New York / New Jersey
Office: Customer floor
Type: Full-time, founding role
Compensation: Base + equity
Travel: Significant, travel-heavy
Walk into any busy distribution center and you'll see the problem we're solving. Thousands of different products, packaging that changes every season, light that shifts every time a dock door opens. Most picking and case handling is still done by hand, because robots haven't been able to make sense of a scene like that and act on it reliably.
We're hiring the engineer who will lead how our robots see and learn. You'll own perception and robot learning for VisionOS, make the core technical calls, and take our models from simulation onto real robots on customer floors.
About Destro AI
Destro is building the intelligence layer for robotics. MothershipOS coordinates robots, people, and workflows across an entire operation in real time, and it's already running inside large enterprise operations. VisionOS brings that intelligence to the individual robot: perceiving what's in front of it, deciding what to do, and handling it. That's where you come in.
Our founders previously led teams at HAI Robotics, Dexterity, and Matic, and we closed our Seed round at one of the highest valuations for a robotics company at this stage.
Why you'll want this job
- You set the direction. VisionOS has a clear product vision and a foundation in place. You'll own the perception and learning stack and decide how we improve and build it.
- Your work goes live fast. Our customers' warehouses are our lab, so your models get tested in real operations in months, not years.
- You own the whole loop, from data collection and training to deployment on real robots.
- Startup speed. Small team, fast decisions. If you can make the case for something, you can start building it that week.
- You grow as we grow. Your scope expands with every new task, robot, and customer, and you'll work side by side with founders who have scaled robotics before.
- Founding equity and a real stake in where the company goes.
What you'll do
- Perception: build 3D scene understanding (RGB-D fusion, point clouds, segmentation), 6-DOF pose estimation for products and cases, and zero-shot identification with vision-language models.
- VLA models: make vision-language-action models the backbone of how our robots carry out tasks. Run pre-training and fine-tuning, including synthetic data in simulation and LoRA or full fine-tunes on GPUs.
- Deployment: get models running on real robots within tight latency and compute limits, and build the active learning loop that turns mistakes on the floor into better models.
- Data and teleoperation: set up teleoperation, run data collection, and build the pipeline from raw recordings to clean, versioned training data.
For us, a model is done when it works on the floor, not when the eval curve looks good.
Your first month
Get to know our robots, simulation environment, and customer floors, then write the perception and learning plan you want to bet on. By end of the month, you'll have a first model running on a physical robot doing a real task.
What you bring
- 3 to 5 years of hands-on work in robotics, computer vision, or ML, ideally on physical robots
- Hands-on model training (pre-training or fine-tuning on GPUs) and experience shipping models to production inference under latency constraints
- 3D perception experience with point clouds and depth cameras
- Experience deploying on real robot hardware, and comfort working in simulation (MuJoCo, Isaac Sim, or similar)
- Strong Python and C++ (or a similar systems language)
- A habit of getting physical prototypes working in weeks, not months, and owning decisions independently
Nice to have
- VLA or manipulation policy training
- Zero-shot pose estimation or segmentation
- Synthetic data generation and domain randomization
- Grasp planning, learned or geometric
- Warehouse or logistics automation
Travel
This role is travel-heavy. Our customer warehouses are our lab, so expect to spend significant amount of your time on site, deploying and improving systems on live floors. If you'd rather find problems on the floor than in a slide deck, you'll love it. If you need to be home most weeks, it's probably not the right fit, and we'd rather say so now.