We teach humanoid robots to do real work.
HumanDbot is building the training layer between a general-purpose humanoid and a useful one. We train skills in simulation first, then sharpen them with data from real buildings, so a robot can take on the repetitive, physical jobs that keep facilities running.
- 01Describe the taskprompt
- 02Generate it in simulationCosmos 3
- 03Test it against real physicsIsaac Sim
- 04Refine and deploy to the robotpost-train
- 05Learn from real deploymentsfield data
Humanoids are arriving. Most can't do a day's work yet.
Robot hardware has improved fast. Teaching a robot a specific job in a specific building is still slow, expensive and specialised.
Every skill needs a team
Today, each new task means ML engineers, a robot on site and weeks of hand-collected demonstrations.
The last mile breaks
Policies trained only in simulation struggle with real friction, clutter and edge cases once deployed.
Labour gaps keep growing
Cleaning, patrol, inspection and care work are hard to staff, and the need repeats every single day.
One platform to train, deploy and improve robot skills
Teleoperation Studio
Operators demonstrate a task remotely. Each demonstration becomes training data, with no ML expertise needed.
Simulation-first engine
Synthetic video and physics simulation bootstrap a skill before a robot ever touches the task. Real data then closes the gap.
Skills library
Trained skills are hardware-agnostic. A skill built once can be adapted to different humanoid brands and sites.
Customers don't buy robots or hire AI teams. They subscribe to a robot that does the job, delivered as Robotics-as-a-Service.
What we believe
Simulate first, deploy second
Simulation is the cheapest place to make mistakes. We test there exhaustively before anything moves in the real world.
Real data is the moat
Open tools can build a pipeline. What compounds is data from robots doing real tasks in real buildings.
Any robot, not one robot
We build for the humanoid market as it is: many manufacturers, fast-changing hardware and no single winner yet.
Built in India, for India first
Our first sites, economics and pricing are designed around Indian facilities and Indian budgets.
Starting with buildings, built for six industries
We begin with facility management, where tasks are repetitive, well-defined and needed daily. The same skills then carry into neighbouring industries.
Praveen Thota
Founder · HyderabadPraveen has spent more than 20 years leading enterprise technology programs, most recently owning AI and agentic-AI products for large-scale software modernization. He started HumanDbot after seeing how much of the work that keeps buildings running still depends on people doing the same physical tasks every day.
- Two decades of enterprise technology leadership, from strategy to go-to-market
- Product owner for AI and agentic-AI platforms at enterprise scale
- Hands-on with NVIDIA Cosmos 3 and Isaac Sim, building the HumanDbot training pipeline himself
From idea to first simulated patrols
- May 2026
Founded as CleanBot Intelligence. Started with a focus on training humanoids for building cleaning and maintenance.
- Jun 2026
Widened the vision. Confirmed that one training approach can serve many robot brands, and expanded beyond facility management.
- Jul 2026
Became HumanDbot. Generated our first synthetic training videos with NVIDIA Cosmos 3 and designed the full training pipeline.
- Aug–Oct 2026
Into physics simulation. Began validating robot movement against real physics in NVIDIA Isaac Sim.
- Now
Preparing first pilots with prospective customers, to start collecting real deployment data.
The robots are coming. We're making sure they know how to work.