Skip to content
View obiedeh's full-sized avatar

Highlights

  • Pro

Block or report obiedeh

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
obiedeh/README.md

Obinna Edeh

I design and build Physical AI and edge AI systems, and prove them on real robots and hardware.

I close the gap between a robot that works in the lab and a fleet that works in the field.

I help teams take hardware and AI from prototype to reliable operation at scale. I have done that across deployment and integration programs at AT&T, Google, Amdocs and T-Mobile, leading engineering teams, field crews and vendors through thousands of field sites. One lesson held on every program: a spec sheet describes ideal conditions; the site shows what the hardware actually does.

Physical AI has the same gap. At EmbodiedEdge, my Physical AI lab, I test robots and edge AI on my own hardware and measure what survives the move from benchmark to field.

Open to senior engineering leadership roles in Physical AI, robotics and edge AI.


How I Lead the Work

  • Validate against the site, not the spec sheet.
  • Set the budget before the test.
  • No claim without its evidence.
  • Corrections stay in the record.
  • Humans stay in the loop.

Flagship Systems

Each project links its proof: the measured result, the hardware and the date.

System What the evidence shows Status
Bench2Field
Repository
Proof: Case study · Report
The benchmark said GO. At default clocks the robot would drop frames. On the rover's Jetson Orin NX the detector alone meets its 33.3 ms deadline (26.9 ms); at the board's default clocks the full frame takes 47.8 ms. With the clocks pinned, the same frame takes 21.1 to 21.5 ms, inside the deadline. v1.0.1 released. v2 in progress.
Physical AI on Jetson
Repository
Proof: Case study
A learned policy running on the robot's own compute. GR00T through TensorRT at 9.8 Hz on Jetson AGX Thor, with a failed evaluation corrected and kept on the record. In simulation, demonstrations that keep mistakes and their corrections took cup placement from 0 to 16 of 50. Active.
Physical AI Safety Observability
Repository
Proof: Case study · Showcase
Full build: Report
Priced the safety loop, then scored it. A vision-language model on the device costs 4.35 s per frame at 66.6 W before the grammar, 2.69 s with it. On 60 labeled simulation pictures, required fields took false gear alarms from 11 to 0, and Cosmos 3 Edge found the worker in 44 of 45 pictures against 29. Functional. Live-camera validation in progress.
Urban Edge Vision Analytics
Repository
Proof: Showcase
Full build: Report
One edge box watching several street cameras. Runs end to end on live cameras; measurement is next. On the Jetson AGX Thor, Cosmos 3 Edge draws rare intersection events on demand for the packs to score, and was tested on predicting the next four seconds of a scene. Functional.

Simulation, synthetic data and mock adapters are labelled wherever they appear and never presented as deployment proof.

Supporting Systems

System What the evidence shows Status
Jetson Edge AI Security
Repository
Proof: Case study
Found a power problem in a runtime setting, not the model. Cut a 54.1 W draw on Jetson AGX Thor to idle, 24.3 W, with zero missed deadlines. Measured.
Field Commissioning Copilot
Repository available on request
Proof: Evidence page
Built so a technician never acts on an unsourced number. Answers commissioning questions from manufacturer documentation and says when no source covers the question. Measured locally.
Wireless signal AI
Four studies
Learned models measured against classical baselines, limits included:
▸ Neural receiver: wins at moderate SNR, falls behind above 15 dB
▸ KPI forecasting: beats a naive baseline, except over holidays
▸ Link intelligence: neural channel estimation wins at low SNR
▸ Edge telemetry: 100 AGVs fit a 20 ms control budget; 120 don't
Simulation and public data.

Technical Stack

  • Physical AI and robotics: ROS 2, MoveIt 2, Isaac Sim, Isaac Lab, OpenUSD, GR00T, LeRobot, SLAM
  • World models: NVIDIA Cosmos Reason 2, Cosmos 3 Edge, Cosmos Transfer
  • Edge inference: NVIDIA Jetson AGX Thor and Orin NX, TensorRT, ONNX Runtime, CUDA, vLLM
  • ML and operational AI: Python, PyTorch, vision-language models, retrieval-grounded copilots, evaluation harnesses
  • Infrastructure: Docker, Kubernetes, CI/CD, AWS, Azure, GCP

Contact

Newsletter: At the Edge with Obyno · obiedeh@gmail.com · LinkedIn · embodiededge.ai · obiedeh.github.io

Pinned Loading

  1. bench2field bench2field Public

    Measures how much of a model optimization survives contact with the robot: full-frame latency at the camera's real rate on Jetson Orin NX, not model-only benchmarks. An EmbodiedEdge project.

    Python 1

  2. physical-ai-jetson-robotics physical-ai-jetson-robotics Public

    Physical AI on the robot's own compute: GR00T through TensorRT on Jetson AGX Thor, Isaac Sim evaluation of a 6-DoF arm, ROS 2 on Orin NX, and a correction ledger kept on record. An EmbodiedEdge pro…

    Python 3

  3. physical-ai-safety-observability physical-ai-safety-observability Public

    A safety layer that measures itself before it judges a robot: vision-language safety events on Jetson AGX Thor with Cosmos-Reason2 and Cosmos 3 Edge, operator review, and measured runtime cost. An …

    Python 1

  4. jetson-edge-ai-security jetson-edge-ai-security Public

    Defensive edge telemetry for Jetson nodes with operator-reviewed alerts. Traced a 54.1 W runtime power draw to the inference thread pool and brought it down to idle, 24.3 W. An EmbodiedEdge project.

    Python 1

  5. urban-edge-vision-analytics urban-edge-vision-analytics Public

    One edge box watching several street cameras: stop-sign, speed and moving-object event packs per camera, with operator review. Runs end to end on live cameras. An EmbodiedEdge project.

    Python 1

  6. ai-phy-neural-receiver-benchmark ai-phy-neural-receiver-benchmark Public

    AI-PHY neural receiver benchmark comparing DeepRx-style and LS+LMMSE receivers on a Sionna-modeled 5G NR link.

    Python 3