Backpack: lessons
The onboard computer that runs perception models and policies on time. Heavy inference is handled by the edge server, and servo control and safety stops by the motor drivers.
On the robot
No lessons are published yet. Here are the planned topics.
Overview
The role of the onboard computer, what runs on the board and what stays off it
-
The role of the onboard computer: what runs on the board and what stays off itComing soon
The SoC architecture, the layers of the JetPack stack, and why servos and the safety stop are kept off the board.
Basics
Setting up a Jetson environment (JetPack, power modes, containers), building a TensorRT engine and measuring its latency distribution, running a policy on the board
-
Setting up a Jetson developer kit: JetPack, power modes and containersComing soon
What each JetPack version bundles, nvpmodel and tegrastats, and running ROS 2 in an L4T container.
-
Building a TensorRT engine and measuring its latency distribution: trtexec, FP16, p50 and p99Coming soon
From ONNX to an FP16 engine with trtexec, p50 and p99 latency, and why an engine is tied to the device and software versions it was built on.
-
Running a trained policy on the board: why the robot stutters in a synchronous inference loopComing soon
Per-stage latency measurement and the pause that appears at every action-chunk boundary.
Advanced
Performance loss during long runs: heat, throttling, power modes, long-duration tests
-
Why the robot slows down after 20 minutes: heat, throttling, power modes and long-duration testingComing soon
Thermal throttling and sustained performance, and the trade-off between power modes and heat dissipation.
In practice
Finding the causes of latency, analyzing the GPU timeline with Nsight Systems
-
Finding the source of latency: analyzing the GPU timeline with Nsight SystemsComing soon
Spotting time lost to preprocessing, memory copies, synchronization waits and GPU contention, then confirming the improvement with p99 latency.