Walker is
the syllabus.

A 5′6″ humanoid engineered across mechanics, locomotion, and edge intelligence—and the public record of making the whole system real.

Ownership · designed, funded, and built by one person.

Build status · all 22 actuators ordered Sep 01; physical assembly next.

Mechanical integration / rev. 0622-actuator architecture

1.6764 m: floor → torso
35.4 kg: physics model
22: RobStride motors
8 GB: edge envelope

Independent full-stack build

One builder owns every boundary.

01BodyCAD + actuators
02MotionRL policies
03BrainJetson inference
04Groundingpapers + lectures

Two contiguous 30-day series

The build, split at the neck.

One series earns the compute budget. The other earns the gait. They are separate daily practices pointed at the same robot.

01 / 30 days

Jetson edge inference

Make 8 GB feel intentional.

Thirty measured decisions for the machine above: memory fit, CUDA kernels, TensorRT, quantization, live vision, local language models, and the two clocks between a policy and its motors.

Latency measurement chart showing p50 17.4 milliseconds and p99 18.8 milliseconds — Day 20 / live pipelinep99 is the robot.

Kernel2.34×

Three preprocess kernels fused into one: 479.9 → 205.5 μs.

Pipeline18.8 ms

Measured p99 at 30 FPS with zero dropped frames.

Memory2,442 MB

Left for camera, gait, and margin after the model stack.

Day 7 CUDA fused preprocess benchmark — 07 · Fuse the preprocess

Day 30 RAM budget showing 2442 megabytes remaining — 30 · Defend the RAM budget

The complete working index

30 / 30 artifacts prepared

01
There is no VRAM on this board
02
Power modes, Super, and why TOPS lie
03
The fit equation: weights + KV + vision
04
Instrumentation is the only legal language
05
System occupancy by subtraction
06
Arithmetic intensity of the camera path
07
A fused Ampere preprocess kernel
08
TensorRT plugins when the graph has a hole
09
Rust as host, CUDA as device
10
A reproducible kernel artifact
11
ONNX contracts and CUDA Graphs
12
FP16 baseline, INT8 on trial
13
Name the runtime correctly
14
Quantization Orin can execute
15
Builder memory vs. runtime memory
16
Keep the frame on the GPU side
17
Embeddings as the cheap semantic layer
18
Detectors under a RAM cap
19
Vision tokens are context tokens
20
The live pipeline exam
21
Choose the house SLM
22
Prefill, decode, and the wrong headline
23
Tool calling as the acceptance test
24
Serving shape: process, not notebook
25
Packing STT + SLM + TTS + eyes
26
VLM latency is not VLA latency
27
The Nano-legal VLA menu
28
Architecture that buys Hertz
29
Policy Hz vs. control Hz
30
The budget I would defend in a room

02 / 30 days

RL walking policies

Teach the body to recover.

The hand-written gait proves the model can move. This series begins where it stops: learned locomotion, domain-randomized dynamics, disturbance recovery, and transfer to twenty-two real motors.

Sequential MuJoCo simulation frames showing Walker completing a walking rollout — Pre-policy baseline / MuJoCo40 steps · no fall

Baseline distance6.217 m

36 seconds · 70.6 steps/min · feet-only floor contact

What the series has to answer

01Can one policy tolerate mass and inertia mismatch?
02Can recovery be trained, not hand-scripted?
03Which behaviors survive the sim-to-real crossing?

Walking telemetry dashboard with displacement, center of mass, torso attitude, foot contacts and motor torque — Controller as measurement instrument0 applied torque-envelope violations

Orthogonal track / 28 papers · 15 lectures

Study that outlives this build.

The lectures and paper implementations are not filler between robot updates. They are a separate record of the ideas underneath the work.

15 lecture videos

Back to first principles.

Stanford CS229Learning systems
Stanford CS231nVisual recognition
MIT UnderactuatedDynamics + control
MIT 6.S191Deep learning

28 paper implementations

One builder owns every boundary.

The build, split at the neck.

Back to first principles.

Rebuild the argument.