Mayabazar is building physical environments-as-a-service for embodied AI. You send a scene spec; a film-trained crew in India builds the set overnight; teleoperators generate expert episodes; your policy runs on GPUs inside the building; a QC'd dataset and digital twin come back.
Mountain View, CA · build operations in India
Per-episode cost of teleoperated demonstration data, 100-episode run. Western figures from published market rates; ours from our India cost structure.
| Per episode | Western market | Mayabazar |
| Teleop labor | $9.05 | $1.20 |
| Scene build & change-over | $5.00 | $0.40 |
| QC review | $2.00 | $0.40 |
| Total | ~$16 | ~$2 |
Research-grade estimates. Western teleop market: ~$136/operator-hour at 5–50 episodes/hour (published 2025 figures). Get your scene's instant quote ›
Talk to any robot-learning team and the same three walls come up. The numbers below are the industry's, not ours.
Teleop data runs $136–340 per operator-hour in Western markets. Worse: past a few hundred demos of one task, more of the same adds almost nothing — the value is in new tasks and new scenes. But every new scene is a construction project, so labs keep re-recording the same countertop.
Environment buildout is the unglamorous bottleneck: sourcing, carpentry, setup, teardown. Our scene system turns it into an overnight assembly job — and the second kitchen costs a fraction of the first.
Skilled operators rarely fail, so policies never learn recovery. And most data comes back as an opaque tarball. Our scene-specs carry failure-injection quotas, and every episode ships with an accept/reject QC verdict you can audit.
Designed around one loop: your config in the evening, verified data in the morning.
A YAML config describes the environment, objects, tasks, and acceptance criteria. Simulate one now.
Film-industry art crews construct the physical scene to spec — the craft India's cinema does best.
Trained operators generate expert demonstrations on real hardware in the built scene.
Your checkpoints run on GPUs inside the facility. Weights never leave the building.
Accepted episodes, a QC report, and a digital twin of the exact scene, delivered.
Every deliverable ships with the evidence attached.
LeRobot-format episodes, each passed through automated QC:
A simulation-ready model of the exact physical scene your data came from — geometry, materials, object poses — so sim-to-real starts aligned.
A hash-chained ledger records every episode, operator session, and QC decision. Safety monitors (force limits, geofence) log every intervention.
Mayabazar (1957) is Indian cinema's masterclass in constructing worlds that feel real. That craft — art departments that build a convincing kitchen, clinic, or warehouse in a night — is exactly the bottleneck in scaling robot data collection. India has the world's deepest bench of it, plus the operator talent pool and cost structure to make per-episode economics work.
Art crews routinely stand up production-quality environments overnight. We point that muscle at robot training scenes.
Configs pushed from the US evening land at the start of an India workday. Data returns before your morning standup.
Customer policies run on-prem, inside the building. Your model weights and task specs stay in a controlled perimeter.
We start where the research community already lives: ALOHA-class bimanual stations with teleoperation built in, emitting LeRobot-format episodes your existing pipeline ingests on day one. Scene-specs are embodiment-agnostic — low-cost arm lanes and industrial platforms get added when a customer's target embodiment requires them, priced into that sprint.
We're selecting a small number of design partners for paid feasibility sprints — the first funded builds of the cell.
Starts with a discovery call and a non-binding letter of intent. You get first-partner pricing and direct input into the scene-spec format.
Honest status: we are pre-buildout. Sprints are scheduled against the first cell, funded by mobilization fees. Early partners know exactly what stage we're at — that's the deal.