Writings / field volume 01

August 2026

From 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.

Side-by-side simulation frames showing one Walker policy falling and another recovering from a 120 newton chest push
Plate 00 / push recoverys105 falls. s108 recovers.

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

Current Walker V1 humanoid CAD render
Object 01 / current CAD / rev. Aug 10

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.

Part one / build record

The machine

Reverse engineering, control, inference, money, and whatever broke next.

01

Already true

The parts are not here yet. These stories are.

RECORDED8 notes · ready now
  1. 01

    The DIP switch that ate the bring-up

    Reverse-framing an undocumented CH340 protocol at 921600 baud.

  2. 02

    The adapter that never worked

    When to stop debugging a part and swap it.

  3. 03

    Position control without the jump

    A four-line safe-start recipe for a 60 N·m motor.

  4. 04

    A cube built from six prisms

    Where agent-made CAD stopped being trustworthy without hand verification.

  5. 05

    Pricing a humanoid honestly

    How a $6.8k estimate became a $10.9k–$17.9k audit.

  6. 06

    One arm

    The reach physics and budget behind an asymmetric robot.

  7. 07

    Three days of almost buying a Unitree

    The build-versus-buy wobble, written down.

  8. 08

    Three walking policies and no robot

    What standing in simulation proves—and what it cannot.

02

Bring-up

Commissioning, control loops, asymmetry, and every fall worth keeping.

ON HARDWARE8 notes · when the boxes open
  1. 09

    First stand

    Restraint rig, precharge, current limits, e-stop chain, and the first load trace.

  2. 10

    Twelve motors, one bus

    Addressing, termination, and what actually saturates first at 1 Mbps.

  3. 11

    PREEMPT_RT, measured

    Stock kernel versus RT patches, answered with a jitter histogram.

  4. 12

    The 50 Hz contract

    IMU read to torque command, including the UDP hop and missed packets.

  5. 13

    One limb, two protocols

    CAN joints and a serial elbow sharing one control loop.

  6. 14

    The sim-to-real ledger

    One row per mismatch: what sim said, what the robot did, what closed it.

  7. 15

    Falls

    Every fall on camera, paired with the log line that explains it.

  8. 16

    Asymmetric

    The gait cost of one powered arm and one mass-mirrored steel ballast.

03

The Jetson notes

Latency, memory, watts, and the compromises hidden by averages.

EDGE LANE6 notes · edge lane
  1. 17

    Day one numbers

    Unoptimized SmolVLA latency, memory, and watts on the Orin Nano 8 GB.

  2. 18

    Quantization against tasks, not perplexity

    INT8 and INT4 on the arm task, including where behavior breaks first.

  3. 19

    TensorRT surgery notes

    What compiled, what fell back, and the kernel that owned the profile.

  4. 20

    Does it all fit

    Gait, VLA, and a talk model sharing 8 GB. Either answer is worth publishing.

  5. 21

    JetPack breaks, fixes follow

    A live ledger of each breakage, fix, issue, and upstreamable patch.

  6. 22

    Where the 320 KB brain lives

    Pi 5 versus Jetson placement, with the UDP tax included.

04

The seam

Where manipulation, balance, and shared silicon start arguing.

SHARED BODY2 notes · shared body
  1. 23

    Arm versus balance

    Center-of-mass disturbance when the arm moves at speed.

  2. 24

    When the VLA starves the gait loop

    Contention, latency budgets, and scheduling on shared silicon.

05

The honest ones

Reviews, re-derivations, money, and the final grade.

FINAL GRADE4 notes · no victory lap
  1. 25

    The review log

    What maintainers taught me, including rejected and rewritten PRs.

  2. 26

    Penance

    Re-deriving agent-written work by hand and naming what I missed.

  3. 27

    What the money bought

    Final BOM against plan, every descope, and what each one cost later.

  4. 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.

01

The deep learning spine

5 papers in this reading cluster.

READING SET5 papers / 9.5 h
  1. 01The mechanism inside every VLA in the stack.Attention Is All You Need (opens in a new tab)2.5 h
  2. 02The residual stream inside the vision tower and policy net.Deep Residual Learning for Image Recognition (opens in a new tab)1.5 h
  3. 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
  4. 04The language-grounded vision ancestry of the VL in VLA.CLIP (opens in a new tab)2.5 h
  5. 05The adaptation method that can actually fit inside 8 GB.LoRA (opens in a new tab)1.5 h
02

RL core

4 papers in this reading cluster.

READING SET4 papers / 6.5 h
  1. 06The opening move every modern deep-RL conversation assumes.Playing Atari with Deep Reinforcement Learning (opens in a new tab)1 h
  2. 07The exact algorithm training the gait in rsl_rl.Proximal Policy Optimization Algorithms (opens in a new tab)1.5 h
  3. 08The depth hiding behind the sentence ‘I used PPO.’Generalized Advantage Estimation (opens in a new tab)2 h
  4. 09The off-policy base of the 15-minute humanoid result.Soft Actor-Critic (opens in a new tab)2 h
03

Locomotion and sim-to-real

7 papers in this reading cluster.

READING SET7 papers / 10 h
  1. 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
  2. 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
  3. 12The sim-to-real primitive behind every randomized range.Domain Randomization for Transferring Deep Neural Networks (opens in a new tab)1 h
  4. 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
  5. 14The closest published ancestor of the walking demo.Real-World Humanoid Locomotion with Reinforcement Learning (opens in a new tab)1.5 h
  6. 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
  7. 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
04

Imitation and manipulation

5 papers in this reading cluster.

READING SET5 papers / 10.5 h
  1. 17The canonical repair for compounding error in imitation.DAgger (opens in a new tab)2 h
  2. 18The beginning of the robotics-transformer line at scale.RT-1 (opens in a new tab)1.5 h
  3. 19The low-cost action-chunking recipe behind the arm lane.ACT and ALOHA (opens in a new tab)1.5 h
  4. 20Why generative action heads became a serious default.Diffusion Policy (opens in a new tab)2.5 h
  5. 21The math under π0’s action head—and the hardest read here.Flow Matching for Generative Modeling (opens in a new tab)3 h
05

VLAs and robot foundation models

8 papers in this reading cluster.

READING SET8 papers / 16 h
  1. 22Where web knowledge transferring into robot control became the claim.RT-2 (opens in a new tab)1.5 h
  2. 23The language for a Walker-specific embodiment story.Open X-Embodiment (opens in a new tab)1.5 h
  3. 24The open 7B baseline whose edge tradeoffs resemble the roadmap.OpenVLA (opens in a new tab)2 h
  4. 25A VLM plus flow head: the model family the build has to confront.π0 (opens in a new tab)2.5 h
  5. 26The public direction for generalization beyond familiar rooms.π0.5 (opens in a new tab)2 h
  6. 27RL on top of a VLA; where the PPO depth pays off.π*0.6 / RECAP (opens in a new tab)2.5 h
  7. 28The small VLA intended to run on the actual arm.SmolVLA (opens in a new tab)2 h
  8. 29A humanoid foundation model with system 1/system 2 seams to test.GR00T N1 (opens in a new tab)2 h
06

World models and self-supervision

5 papers in this reading cluster.

READING SET5 papers / 10.5 h
  1. 30The inexpensive origin of learning inside a dream.World Models (opens in a new tab)1 h
  2. 31The standard test for imagination-based RL across domains.DreamerV3 (opens in a new tab)2 h
  3. 32Predict in latent space, not pixels, in its cleanest form.I-JEPA (opens in a new tab)1.5 h
  4. 33A video world model with robot-planning results attached.V-JEPA 2 (opens in a new tab)3.5 h
  5. 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
07

Edge efficiency and kernels

4 papers in this reading cluster.

READING SET4 papers / 6.5 h
  1. 35The intellectual frame for small-from-big at the edge.Distilling the Knowledge in a Neural Network (opens in a new tab)1 h
  2. 36Why naïve 8-bit inference breaks and where outliers live.LLM.int8() (opens in a new tab)1.5 h
  3. 37The 4-bit method closest to the planned Orin deployment.AWQ (opens in a new tab)1.5 h
  4. 38The IO-aware kernel mindset behind honest profiling.FlashAttention (opens in a new tab)2.5 h
08

The VLM recipe and the GEAR angle

2 papers in this reading cluster.

READING SET2 papers / 3 h
  1. 39The recipe under the VLM, and therefore under the VLA.Visual Instruction Tuning (opens in a new tab)1.5 h
  2. 40A live experiment: let an LLM rewrite the walking reward.Eureka (opens in a new tab)1.5 h
68

End note

Write it because it happened and the numbers are real.
Walker V1 / 68 working threads / Aug 2026