About Stealth Startup
It is a pre-seed robotics and AI startup building custom software for off-the-shelf humanoid robots to operate autonomously in construction environments. Founded by Benedict Floyd-Sanchez, who brings 5 years of construction industry experience, the company is developing a “world model” — intelligent software that enables robots to think, plan, and act in the chaotic, ever-changing reality of a live construction site. Backed by Anarock (Indian property tech), the company is currently at the MVP stage with a virtual demo and raising capital to build out its engineering team.
The Mission
You will design the cognitive architecture that lets our robot think, plan, and act over long horizons in an environment that never stays the same for more than five minutes. Full stack from task-level planning down to policy execution, ensuring our system generalises from simulation into the chaotic reality of a live construction site — moving workers, unpredictable debris, and tools that land nowhere near where they were dropped.
What You’ll Own
- Design and ship a hierarchical planning system capable of multi-step construction tasks under partial observability (POMDPs, foundation-model planners)
- Drive simulation-to-real transfer strategy across Isaac Sim, MuJoCo, and physical hardware — close the gap relentlessly
- Architect transformer-based policies for long-horizon manipulation (tile laying, rebar threading, drywall handling)
- Integrate neuro-symbolic reasoning so the robot can represent and reason over structured site constraints (safety zones, load tolerances, sequencing rules)
- Build real-time adaptation loops that handle dynamic obstacles — moving scaffolding, other workers, tool drops — without stopping
- Define the evaluation regime: what ‘good’ looks like in sim and on-site, and how to measure the gap
- Develop the “world model” software powering vision language models and reinforcement learning for autonomous construction robotics
Required Tech Stack
- Python (primary language) + C++
- Isaac Sim & MuJoCo (robotics simulators) — large-scale training runs
- SLAM (Simultaneous Localisation and Mapping)
- LangGraph / agent workflow frameworks
- JAX or PyTorch at scale
- Vector databases / Knowledge Graph (GraphDB)
- Shipped a policy that runs on a physical humanoid or dexterous manipulator — not just in sim
- Deep expertise in hierarchical reinforcement learning, model-based planning, or foundation models for robotics
- Hands-on with Isaac Sim, MuJoCo, or PyBullet for large-scale training runs
- Fluent in Python + C++ comfortable with JAX or PyTorch at scale
- Strong understanding of generative AI, prompt engineering, and SaaS AI ecosystem workflows
- Strong communicator — able to translate cutting-edge research into decisions the whole team can act on
- High agency, self-starter — can take user stories/architecture and deliver independently with minimal hand-holding
Strong Bonus (Nice-to-Have)
- Published work on sim-to-real transfer, meta-learning, or robot foundation models
- Experience in safety-critical deployments (construction, industrial, medical)
- Familiarity with RT-2, SayCan, or similar vision-language-action architectures
- PhD in robotics, ML, or a related field (experience is the primary filter)
- Team management skills for potential future leadership responsibilities