Skip to content

Harden hackathon Rerun, Viewer, and SmolVLA workflow - #8

Merged
toyeshhm merged 7 commits into
mainfrom
codex/hackathon-readiness
Jul 19, 2026
Merged

Harden hackathon Rerun, Viewer, and SmolVLA workflow#8
toyeshhm merged 7 commits into
mainfrom
codex/hackathon-readiness

Conversation

@toyeshhm

Copy link
Copy Markdown
Collaborator

What this integrates

  • merges the SmolVLA warm-start and Rerun Query API branches onto current main
  • logs the vendored reBot URDF as real meshes/transforms with a safe text fallback
  • adds a central Rerun Viewer Blueprint for both cameras, 3D robot, goal/position traces, and timeline
  • fixes macOS Rerun native binary discovery for recording verification and Viewer launch
  • documents the tested SmolVLA dependency recipe without downgrading Rerun 0.34
  • adds a truthful hackathon pitch/demo package with synthetic-vs-live evidence boundaries

Verification

  • 17 passed: Rerun hardening + Viewer/library tests
  • 151 passed: Person 4 harness/episode/evaluation tests
  • 36 passed: rollout contract and safety tests
  • real 20-step public SmolVLA MPS fine-tune completed, checkpoint saved and freshly reloaded
  • fresh synthetic record -> Query API -> native LeRobot export -> fresh dataset load -> fake replay completed

Physical boundary

No reBot serial hardware is connected to this Mac. Physical motion, live can sorting, physical replay, and learned-policy success are not claimed. Those remain gated by reviewed calibration, live hardware/camera checks, a clear workspace, and a dedicated e-stop operator.

csplatti and others added 7 commits July 18, 2026 18:27
Document the verified local SmolVLA training setup (Python 3.11 venv,
vendored lerobot, the mandatory --dataset.video_backend=pyav workaround)
and git-ignore outputs/ so training artifacts aren't committed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Fine-tune the pretrained lerobot/smolvla_base (--policy.path) rather than
--policy.type=smolvla, which trains from scratch and won't learn a usable
policy on ~150 episodes. Matches the official "warm-start from SO-101 base"
recipe.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
smolvla_base expects cameras camera1/2/3; our data is front/side. Add the
required --rename_map and --policy.empty_cameras=1 to the train + smoke-test
commands. Full warm-start recipe verified end-to-end on MPS.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings July 19, 2026 06:46
@toyeshhm
toyeshhm merged commit 6ae546e into main Jul 19, 2026
1 check passed
@toyeshhm
toyeshhm deleted the codex/hackathon-readiness branch July 19, 2026 06:47

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

This PR hardens the reBot Rerun/Query API workflow and adds a consistent operator-facing Viewer layout, while documenting a verified local SmolVLA fine-tuning recipe and clarifying hackathon demo evidence boundaries.

Changes:

  • Add a purpose-built Rerun Viewer Blueprint and activate it during interactive recordings (while keeping --no-viewer fully headless).
  • Log the vendored reBot URDF as actual meshes/transforms when possible, with a safe text fallback.
  • Document SmolVLA warm-start training steps (MPS + PyAV) and add supporting docs/demo materials; update dependencies and ignore runtime artifacts.

Reviewed changes

Copilot reviewed 14 out of 16 changed files in this pull request and generated 2 comments.

Show a summary per file
File Description
SMOLVLA_TRAINING.md New step-by-step local SmolVLA fine-tuning guide (warm-start, PyAV, camera rename/padding, MPS).
requirements.txt Adds rerun-loader-urdf dependency for URDF mesh logging.
rebot_operator_kit/teleop_gui/training_workspace.py Improves macOS Rerun native binary discovery (CLI vs app bundle).
rebot_operator_kit/teleop_gui/controlled_record.py Improves macOS Rerun native binary discovery for verification and viewer usage.
p5_rerun_port/viewer_blueprint.py New central Viewer Blueprint for cameras, URDF, metrics, and timeline.
p5_rerun_port/urdf_log.py Logs URDF meshes/transforms via rerun-loader-urdf with fallback to text path.
p5_rerun_port/tests/test_rerun_hardening.py New tests covering vendored URDF resolution/logging, Query API smoke, and blueprint/headless behavior.
p5_rerun_port/takes.py Adds optional Viewer Blueprint initialization and activation to recording setup.
p5_rerun_port/README.md Adds link to hackathon demo doc and notes Blueprint-enabled interactive recordings.
p5_rerun_port/constants.py Prefers repo-vendored URDF path as first resolution candidate.
docs/TEAM.md Updates SmolVLA guidance to point to the verified warm-start recipe.
docs/Rerun_bounty_progress.md Marks Viewer Blueprint slice as done (synthetic) and updates bounty status notes.
docs/p5_rerun_port/README.md Documents Viewer Blueprint workflow and confirms headless --no-viewer behavior.
docs/p5_rerun_port/examples/hackathon_smoke_query_report.md New synthetic Query API report example with explicit evidence boundary.
docs/HACKATHON_SUBMISSION_DEMO.md New hackathon pitch/demo script with explicit synthetic vs live claims and checklist.
.gitignore Ignores .venv-lerobot/ and outputs/ artifacts.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment thread p5_rerun_port/takes.py
except Exception:
pass
rec.set_sinks(*sinks)
rec.set_sinks(*sinks, default_blueprint=blueprint)
@@ -40,6 +40,14 @@

# Prefer SDK URDF; fall back to env override.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants