Skip to content
See the work

The field-readiness lab for Physical AI

Physical AI that works outside the demo and in the real world.

EmbodiedEdge tests, integrates and measures robots and edge AI systems on real hardware, in real conditions, and publishes what it finds.

See the work See the evidence
The lab's robot arm at the same Ludo board, in Isaac Sim and in the real world: real photoThe lab's robot arm at the same Ludo board, in Isaac Sim and in the real world: simulation renderSimulation · Isaac SimReal world · Synria arm

Sim ↔ real Same arm, same board: the Isaac Sim digital twin and the physical robot on the bench, in turn. Click to switch.

  • Built onHardware integration at scale
  • BuildingPhysical AI and robotics systems
  • OperatingEdge deployment and observability

What we do

Test. Integrate. Deploy. Operate.

A policy that passes in simulation meets sensor noise, thermal limits, latency budgets, and operators who have to trust it. Most Physical AI stalls right there. This is the method the lab works through, and each stage ends with something that can be inspected.

  1. 01

    Test Measured

    Sim-to-real evaluation in a digital twin, then benchmarking on the real hardware: latency, power, thermals, and failure modes.

    • Isaac Sim
    • OpenUSD
    • TensorRT
    • Power and thermal
    OutputA dated measurement report
  2. 02

    Integrate Active

    ROS 2 integration, perception and inference on Jetson, and sensor and actuator bring-up, instrumented from day one.

    • ROS 2
    • Jetson Thor / Orin
    • ONNX
    • Sensors
    OutputA working, instrumented system
  3. 03

    Deploy From prior programs

    How a system should reach the field: installed, commissioned, and verified on site, with the sequence written down before anyone drives out to do it.

    • Rollout sequence
    • Commissioning
    • Runtime safety
    • Telemetry
    OutputCommissioning checks others can repeat
  4. 04

    Operate Planned

    How a system should be kept running after launch: monitoring, maintenance, and updates, with telemetry an operator can act on and a clear line on what stays a human decision.

    • Monitoring
    • Maintenance
    • Updates
    • Operator review
    OutputTelemetry an operator can act on

Where it applies

Orange telemetry gradient

Current proof points

Built to be inspected.

  • AGX Thor101.6 ms TRT · 9.8 Hz2026-08-20
  • Thermals42.0 °C peakAGX Thor · 2026-08-20
  • Isaac Sim35/36 sim turns · 19.2 mm p50RTX 5090 · 2026-08-20
  • Edge IDS0.0237 ms p95 · 54.1 W → 24.3 WAGX Thor · 2026-09-08
Hardware
Jetson AGX Thor
Workload
Intrusion detector on live traffic, ONNX Runtime
Measured
Board power, junction temperature, p95 latency
What failed
Default threading drew 31 W more for the same work
What changed
Detector pinned to a single thread
Result
54.1 W → 24.3 W, about 15 °C cooler, held across three sessions. Worst single event still unresolved.

One study, the whole way through

Every result the lab publishes is written in this order: hardware, workload, what was measured, what failed, what changed, and the result, with the open items left in.

Read the study How we measure
Chart: 31 W more power. Same work.
31 W more power. Same work.
Chart: The gap held in all three sessions.
The gap held in all three sessions.
Chart: About 15 °C cooler at the junction.
About 15 °C cooler at the junction.
Chart: Lower power, lower p95 latency.
Lower power, lower p95 latency.
Chart: Where the 31 W goes.
Where the 31 W goes.
Chart: The worst single event is unresolved.
The worst single event is unresolved.

From the bench

Systems, not slides.

All work
Edge inferenceActive
Full frame on the rover
47.8 ms p95
Model alone, same board
26.9 ms
Frames the pipeline sustains
22.9 fps

Bench2Field

Measures how much of a model optimization survives contact with the robot. The model alone met its deadline on the rover; the full frame did not.

Read the build
ManipulationActive

Physical AI on Jetson

Robot manipulation program spanning measured Jetson inference and probes, measured Isaac Sim evaluation, and an implemented but unmeasured Orin ROS 2 host.

Read the build
Safety & observabilityActive

Physical AI Safety Observability

Runtime safety layer on Jetson AGX Thor that measures pipeline performance and model inference, so operators can see when the system is outside safe limits.

Read the build
Edge inferenceActive

Jetson Edge AI Security

Defensive edge telemetry runtime with measured Jetson AGX Thor inference and thread-pool power, plus operator-reviewed alerts.

Read the build
NavigationStarting

Autonomous Navigation on ROSMASTER

SLAM and navigation on a mobile rover with Jetson Orin NX onboard, starting from measured bring-up instead of assumed performance.

Read the build
Edge inferenceActive

Urban Edge Vision Analytics

One edge box watching several street cameras, turning what a vision-language model sees into traffic events an operator reviews. Runs end to end on live cameras.

Read the build
Safety & observabilityActive

Edge Telemetry Lakehouse

One governed data platform for edge AI device events, from raw arrival to operator answers, with every bad record quarantined for a reason. Runs end to end on synthetic replay.

Read the build
Portrait of Obinna Edeh

Behind the lab

Obinna Edeh

Obinna's career has been hardware integration: installing, commissioning, and troubleshooting physical systems and keeping them running on site, including leading a nationwide deployment. That work taught him that the hard part of any system is rarely the design. It is getting the hardware to work where it lives and keeping it working. EmbodiedEdge applies that discipline to robots and edge AI.

The lab runs on its own hardware: Jetson AGX Thor and Orin NX, a Synria arm, a ROSMASTER rover, and an RTX 5090 workstation for simulation. Every number published here was measured on one of them.

About the lab

Contact

Let's build something that can leave the lab.

The lab publishes what it measures. If you are working on the same gap, or want to dig into a result, the inbox is open.