Physical AI · Build in Public
Under construction
VAJRA
Autonomous find-and-fetch robot.
Same governance instincts as Agentify (deny-by-default, budget routers, eval harnesses) re-applied to physical actions.
Jetson Orin NanoROS 2 HumbleNanoOWL
Demo · Coming soon
Tabletop "fetch the red mug" demo
15-second autoplay loop lands here once the v0.1 milestone clip is recorded, actively in build. Tracked in the status checklist below.
Build stack
The exact components and versions in the current build.
Hardware
- Compute
- NVIDIA Jetson Orin Nano Super (8 GB)
- Locomotion
- Differential-drive base (mid-build)
- Manipulation
- 5-DOF arm + parallel gripper (sourced)
- Sensing
- RGB camera, IMU; depth sensor planned
- Compute budget
- ₹35-45K total component spend
Software
- OS / middleware
- Ubuntu 22.04 + ROS 2 Humble (Jazzy migration in progress)
- Perception
- NanoOWL (TensorRT-optimized OWL-ViT), open-vocabulary detection
- Foundation reference
- NVIDIA GR00T N1.7 + Cosmos 3 (April 2026 release)
- Navigation
- Nav2 stack: costmap + behavior tree planner
- Governance
- Capability-lease envelope per action (deny-by-default)
- Telemetry
- OpenTelemetry → Datadog; budget routers per task class
Status
Honest checklist. Anything unchecked is intent, not capability.
3 of 6 complete
- Hardware sourced; Jetson Orin Nano boots.(done)
- NanoOWL TensorRT engine threshold-tuned.(done)
- Object recognition demo on tabletop scene.(done)
- ROS 2 nodes for locomotion + manipulation (in progress).(not done)
- End-to-end "fetch the red mug" demo recording.(not done)
- Public Discord with weekly office-hour build session.(not done)
Build log
Public, dated, append-only on GitHub. No demo cuts, no marketing lift.
Last updated 2026-04-27.