Writings / field volume 01
August 2026From the Walker prototype archive
The robot fell at 120 N.
Then it didn’t.
Sixty-eight entries from the earlier five-foot prototype. The failures stay in because they are the part that teaches.

Same robot, same 120 N chest push. The walking montage means nothing without this series next to it.

Walker V1 / prototype record
The earlier machine these notes answer to.
A five-foot humanoid built from the bus outward. Every number below has already changed a design decision.
- 12 motors
- 8 × RS03, 4 × RS04
- 1 Mbps
- one shared CAN bus
- 50 Hz
- the gait loop must keep
- 8 GB
- for policy and VLA together
Selection ruleA reflection only exists when braided with an implementation or a measurement.
01 / on the bench / 006
Six live threads.
A short index, with the proof left attached.
The DIP switch that ate the bring-up
Reverse-framing the undocumented AT serial protocol behind a silent RS03 and turning it into a custom python-can bus.
921600baud
Position control without the jump
Writing current mechpos into loc_ref before enable. Four lines that remove the jump from a 60 N·m actuator.

Three walking policies and no robot
Training standing and gait in MuJoCo on a laptop CPU, months before any hardware existed.
The spreadsheet, not the vibes.Twelve motors, one bus
Bringing 8 RS03s and 4 RS04s onto one 1 Mbps bus, then measuring what saturates first.
Measured before it became a bottleneck.
Part one / build record
The machine
Reverse engineering, control, inference, money, and whatever broke next.
Already true
The parts are not here yet. These stories are.
- 01
The DIP switch that ate the bring-up
Reverse-framing an undocumented CH340 protocol at 921600 baud.
- 02
The adapter that never worked
When to stop debugging a part and swap it.
- 03
Position control without the jump
A four-line safe-start recipe for a 60 N·m motor.
- 04
A cube built from six prisms
Where agent-made CAD stopped being trustworthy without hand verification.
- 05
Pricing a humanoid honestly
How a $6.8k estimate became a $10.9k–$17.9k audit.
- 06
One arm
The reach physics and budget behind an asymmetric robot.
- 07
Three days of almost buying a Unitree
The build-versus-buy wobble, written down.
- 08
Three walking policies and no robot
What standing in simulation proves—and what it cannot.
Bring-up
Commissioning, control loops, asymmetry, and every fall worth keeping.
- 09
First stand
Restraint rig, precharge, current limits, e-stop chain, and the first load trace.
- 10
Twelve motors, one bus
Addressing, termination, and what actually saturates first at 1 Mbps.
- 11
PREEMPT_RT, measured
Stock kernel versus RT patches, answered with a jitter histogram.
- 12
The 50 Hz contract
IMU read to torque command, including the UDP hop and missed packets.
- 13
One limb, two protocols
CAN joints and a serial elbow sharing one control loop.
- 14
The sim-to-real ledger
One row per mismatch: what sim said, what the robot did, what closed it.
- 15
Falls
Every fall on camera, paired with the log line that explains it.
- 16
Asymmetric
The gait cost of one powered arm and one mass-mirrored steel ballast.
The Jetson notes
Latency, memory, watts, and the compromises hidden by averages.
- 17
Day one numbers
Unoptimized SmolVLA latency, memory, and watts on the Orin Nano 8 GB.
- 18
Quantization against tasks, not perplexity
INT8 and INT4 on the arm task, including where behavior breaks first.
- 19
TensorRT surgery notes
What compiled, what fell back, and the kernel that owned the profile.
- 20
Does it all fit
Gait, VLA, and a talk model sharing 8 GB. Either answer is worth publishing.
- 21
JetPack breaks, fixes follow
A live ledger of each breakage, fix, issue, and upstreamable patch.
- 22
Where the 320 KB brain lives
Pi 5 versus Jetson placement, with the UDP tax included.
The seam
Where manipulation, balance, and shared silicon start arguing.
- 23
Arm versus balance
Center-of-mass disturbance when the arm moves at speed.
- 24
When the VLA starves the gait loop
Contention, latency budgets, and scheduling on shared silicon.
The honest ones
Reviews, re-derivations, money, and the final grade.
- 25
The review log
What maintainers taught me, including rejected and rewritten PRs.
- 26
Penance
Re-deriving agent-written work by hand and naming what I missed.
- 27
What the money bought
Final BOM against plan, every descope, and what each one cost later.
- 28
October 31, graded
The demo commitment measured pass or fail, written whichever way it lands.
Part two / reading record
The literature
Forty papers, ≈73 hours. Not summaries—each one has to answer to the build.
The deep learning spine
5 papers in this reading cluster.
- 01The mechanism inside every VLA in the stack.Attention Is All You Need (opens in a new tab)2.5 h
- 02The residual stream inside the vision tower and policy net.Deep Residual Learning for Image Recognition (opens in a new tab)1.5 h
- 03How images become tokens—and what that costs on the Jetson.An Image is Worth 16×16 Words (opens in a new tab)1.5 h
- 04The language-grounded vision ancestry of the VL in VLA.CLIP (opens in a new tab)2.5 h
- 05The adaptation method that can actually fit inside 8 GB.LoRA (opens in a new tab)1.5 h
RL core
4 papers in this reading cluster.
- 06The opening move every modern deep-RL conversation assumes.Playing Atari with Deep Reinforcement Learning (opens in a new tab)1 h
- 07The exact algorithm training the gait in rsl_rl.Proximal Policy Optimization Algorithms (opens in a new tab)1.5 h
- 08The depth hiding behind the sentence ‘I used PPO.’Generalized Advantage Estimation (opens in a new tab)2 h
- 09The off-policy base of the 15-minute humanoid result.Soft Actor-Critic (opens in a new tab)2 h
Locomotion and sim-to-real
7 papers in this reading cluster.
- 10The result that made learned control on real legs credible.Learning Agile and Dynamic Motor Skills for Legged Robots (opens in a new tab)2 h
- 11The parallel recipe behind the training configuration.Learning to Walk in Minutes Using Massively Parallel Deep RL (opens in a new tab)1.5 h
- 12The sim-to-real primitive behind every randomized range.Domain Randomization for Transferring Deep Neural Networks (opens in a new tab)1 h
- 13The standard answer to surviving reality without per-robot tuning.RMA: Rapid Motor Adaptation for Legged Robots (opens in a new tab)1.5 h
- 14The closest published ancestor of the walking demo.Real-World Humanoid Locomotion with Reinforcement Learning (opens in a new tab)1.5 h
- 15One open mechanical reference among several: useful for proportions, not a blueprint for the integrated system.Berkeley Humanoid Lite (opens in a new tab)1 h
- 16A speed record concrete enough to reproduce and interrogate.Learning Sim-to-Real Humanoid Locomotion in 15 Minutes (opens in a new tab)1.5 h
Imitation and manipulation
5 papers in this reading cluster.
- 17The canonical repair for compounding error in imitation.DAgger (opens in a new tab)2 h
- 18The beginning of the robotics-transformer line at scale.RT-1 (opens in a new tab)1.5 h
- 19The low-cost action-chunking recipe behind the arm lane.ACT and ALOHA (opens in a new tab)1.5 h
- 20Why generative action heads became a serious default.Diffusion Policy (opens in a new tab)2.5 h
- 21The math under π0’s action head—and the hardest read here.Flow Matching for Generative Modeling (opens in a new tab)3 h
VLAs and robot foundation models
8 papers in this reading cluster.
- 22Where web knowledge transferring into robot control became the claim.RT-2 (opens in a new tab)1.5 h
- 23The language for a Walker-specific embodiment story.Open X-Embodiment (opens in a new tab)1.5 h
- 24The open 7B baseline whose edge tradeoffs resemble the roadmap.OpenVLA (opens in a new tab)2 h
- 25A VLM plus flow head: the model family the build has to confront.π0 (opens in a new tab)2.5 h
- 26The public direction for generalization beyond familiar rooms.π0.5 (opens in a new tab)2 h
- 27RL on top of a VLA; where the PPO depth pays off.π*0.6 / RECAP (opens in a new tab)2.5 h
- 28The small VLA intended to run on the actual arm.SmolVLA (opens in a new tab)2 h
- 29A humanoid foundation model with system 1/system 2 seams to test.GR00T N1 (opens in a new tab)2 h
World models and self-supervision
5 papers in this reading cluster.
- 30The inexpensive origin of learning inside a dream.World Models (opens in a new tab)1 h
- 31The standard test for imagination-based RL across domains.DreamerV3 (opens in a new tab)2 h
- 32Predict in latent space, not pixels, in its cleanest form.I-JEPA (opens in a new tab)1.5 h
- 33A video world model with robot-planning results attached.V-JEPA 2 (opens in a new tab)3.5 h
- 34The map for where world-model systems are trying to go.A Path Towards Autonomous Machine Intelligence (opens in a new tab)2.5 h
Edge efficiency and kernels
4 papers in this reading cluster.
- 35The intellectual frame for small-from-big at the edge.Distilling the Knowledge in a Neural Network (opens in a new tab)1 h
- 36Why naïve 8-bit inference breaks and where outliers live.LLM.int8() (opens in a new tab)1.5 h
- 37The 4-bit method closest to the planned Orin deployment.AWQ (opens in a new tab)1.5 h
- 38The IO-aware kernel mindset behind honest profiling.FlashAttention (opens in a new tab)2.5 h
The VLM recipe and the GEAR angle
2 papers in this reading cluster.