Sources and attribution#
The guide connects public code to recorded hardware, simulation and offline results. Choose the implementation version before adapting a procedure: the successful August demo and the later calibration development are separate branches.
Published repositories#
Repository / branch |
Inspected revision |
What belongs here |
|---|---|---|
|
August 25 demo baseline: seated control, cube perception, grasp/planning contracts, stacking, recording and conversion |
|
|
Later calibration runtime, diagnostics, shared-control changes and regression fixtures |
|
|
Reusable calibration tools, bundle support, wrist markers, V7 torso structure and R2 coupons |
|
|
Offline GraspGen-X descriptors, qualification, support studies and assembly planning |
|
|
Initial MuJoCo/OpenHomie backend, supplied on |
|
|
Passive recording, baseline-relative visualization and taxel mapping |
|
|
Printable targets, detector and stateless pose hypotheses used by tabletop |
The tabletop main preserves the runtime from d1b0103, the commit recorded by
the August 25 runs. September carries the September 4 checkpoint and later
development together. The listed checkouts include subsequent README edits and
the AprilCube publication fix below; that fix changes the dependency reference
and provenance entry, with the detector implementation and calibration bundle
unchanged. Code links elsewhere retain the earlier inspected file revisions
(7400aff for the published August baseline, 59c21b1 for September).
Use the setup checkouts for
installation. Neither a newer branch nor a passing offline test replaces the
demonstrated configuration.
The fixture repository remains separate from tabletop. Its consolidated main includes both V7/R2 and the earlier calibration-bundle addition; there is no need to choose between former development branches to obtain those files. It still has the documented physical fit limits.
AprilCube dependency revisions#
Tabletop uses AprilCube b7f86f2 from the existing public fork. Its implementation,
tests and assets match the originally recorded local dependency 9381a52 byte
for byte. It includes the planar-pose disambiguation and stateless pose-hypothesis
API that were missing from the older public fork. The historical calibration
bundle remains unchanged.
The grasp/assembly repository uses AprilCube 020777c, which preserves its
original fc18d50 implementation and assets. It does not need a detector upgrade.
The two dependency references follow the fork’s cleaned history, which excludes
private working notes. Older experiment records retain their original source
identities; use the current checkout instructions to obtain the dependencies.
The target chapter links print files, detector configuration and design documentation at the published tabletop dependency.
Follow the implementation and its evidence#
The code index supplies exact file URLs, hashes and Python symbols. Its JSON map can be consumed by an agent or checked against a clone. All mapped implementation files are public. A chapter using September-specific code labels that boundary explicitly.
Public technical reports live alongside their implementation—for example the grasp descriptor study, September runtime summary and torso fit record. The media catalog records the selected photographs/videos and their review limits; captions distinguish physical footage from rendered replay.
Personal running notes, design proposals and raw investigation records remain private. They informed the explanations, but this guide does not require readers to obtain a private journal to locate code. Raw sessions and full analysis work directories are separate from a source clone. When their findings are summarized, the chapter states the experiment, measured quantity and limits rather than pretending the underlying dataset is included.
What was checked#
The draft’s source reading covered the G1Pilot, grasp-demo, tabletop and calibration-prototype logs, the later investigation ledger and the V7 fit update. Inherited prototype notes were treated as shared history, not independent trials. Later operator corrections govern conflicting older reports. Calibration corrections extend through September 7 and fixture fit evidence through September 13, 2026.
Unless labelled as an independently rerun check, experiment results and historic test counts are reported from retained evidence. During source publication, selected September offline regressions passed 317 tests with seven skipped. That checked software packaging and regressions; it did not commission the final standing lifecycle. The documentation checks verify links, file identities, rendering and the site build without operating the robot.
After the AprilCube publication fix, a fresh GitHub tabletop checkout resolved
its submodules recursively. The read-only inspect command matched the recorded
dependency pins and loaded the saved calibration bundle using the existing
Python environment. Git LFS downloads were disabled for this checkout check;
model materialization and a clean environment installation remain separate checks.
The five August 25 episodes are available as a LeRobot download. The August 12 calibration archive includes original captures, selected fit inputs and the historical bundle. Both have versioned downloads and checksums on Google Drive, separate from the source repositories.
Attribution and continuity#
The research and this guide are by
srinivas (sri299792458), building on Unitree’s
SDK/models, AprilCube, GraspGen-X, CuRobo, Isaac/PhysX, Newton, MuJoCo, OpenHomie,
GR00T, ROS, Pinocchio/OpenSoT and SPARK. Including a dependency does not imply
original authorship of that dependency.
The author’s SPARK documentation provides the reference for practical depth. A later lab copy should preserve attribution, useful public history and access to published source/media. No blanket documentation license has been assigned yet; preserve existing upstream notices and settle the license before the final handoff.