Imported from Mizoreww/agents-for-robotics-talk (
.build/revision_0_14_baseline/AGENTS.md). Install upstream withnpx skills add Mizoreww/agents-for-robotics-talk --skill revision_0_14_baseline. Copyright stays with the author.
Agents for Robotics
One self-contained HTML technical talk — 35 English slides, Chinese speaker script, 18 embedded clips, ~135 MB — plus its research notes and build scripts. Speaker Zimo Huang, talk date 2026-09-16.
Build and verify
Never hand-edit output/Agents_for_Robotics_Self_Contained.html; it is generated. Edit the sources in
.build/ and rebuild. Scripts hard-code the project path /home/limx/Desktop/agent_for_robotics — update
it if the repo moves. Run from the project root:
python3 .build/test_report_figures_v013.py # original report images and quantitative lineage
python3 .build/test_evidence_v011.py # new source identity and query-time gates
python3 .build/test_table_emphasis.py # comparison groups and ties
python3 .build/test_results_v09.py # fail-closed source identity regressions
python3 .build/derive_results_v09.py # validate pinned sources, derive results
python3 .build/build_selfcontained.py # slides, script, HTML, build audit
PYTHONPATH=.build/font_deps python3 .build/build_cjk_font.py # CJK subset from the just-built script
python3 .build/build_selfcontained.py # rebuild to embed the refreshed subset
cp output/Agents_for_Robotics_Self_Contained.html .build/standalone_isolation/index.html
python3 .build/validate_standalone.py --static # static assertions only
python3 -m http.server 8765 --bind 127.0.0.1 --directory .build/standalone_isolation &
NODE_PATH=/home/limx/.cache/codex-runtimes/codex-primary-runtime/dependencies/node/node_modules \
/home/limx/.cache/codex-runtimes/codex-primary-runtime/dependencies/node/bin/node .build/browser_audit.mjs http://127.0.0.1:8765/
python3 .build/validate_standalone.py # full run; binds the audits to the hash
PYTHONPATH=.build/static_render_deps python3 .build/render_static_revision.py # optional static PDF; not part of this revisionThe copy into .build/standalone_isolation/ must precede even --static: the validator asserts that
directory holds exactly one byte-identical file. file:// navigation is blocked by app policy, so the
browser audit runs against loopback (.claude/launch.json, entry deck-isolation). System node is too
old for Playwright — use the bundled modern Node runtime path above. qa_deps, font_deps and static_render_deps
are vendored and excluded from the repo; reinstall with pip install --target (fontTools + brotli +
zopfli, WeasyPrint, defusedxml) and npm i playwright.
The main build interpreter also requires Pillow for original-figure dimensions, not just the
optional PDF renderer. Install it into that interpreter's environment with python3 -m pip install Pillow when pip is available. On this host system Python already provides it; if system pip is
unavailable, the bundled Python at
/home/limx/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/bin/python3
includes Pillow and can replace python3 in the build_selfcontained.py invocations. Installing
Pillow only into static_render_deps does not satisfy the main builder's dependency.
Stack
Python 3.12, bundled Node with Playwright driving /opt/google/chrome/chrome (needed for H.264 playback),
ffmpeg/ffprobe for clip preparation, Noto Sans / Noto Serif / subset Noto Sans CJK SC. Everything is
base64-embedded in the single HTML; CSP sets connect-src 'none', no runtime network requests.
Layout and conventions
.build/— production.build_selfcontained.pydefines slide order, titles and minutes;script_revised.mdis the Chinese-script authority;standalone_shell.html/standalone_player.css/standalone_player.jsare the page shell;standalone_*_audit.jsonrecord the last verified state.output/— deliverables plusSpeaker_Script_Revised.md. The 22-slide PPTX/PDF/ZIP are history and do not match the current page order.research/— notes,sources.json, page snapshots insources/, original figures inassets/.- Slides are English; the script is Chinese, with paper, model and technical terms kept in English.
- Control follows the newly approved sequence: original overview; Hi Robot / Helix / Helix 02 on one page; responsibility-shift diagram; then Astra capability, interface and Direct/Hybrid evidence; summary. Data and Improvement each keep one detailed representative: Agentic Real2Sim and ENPIRE. Community demos stay within chapters, not extra paper deep dives.
- Community demos from X and 小红书 are qualitative: author-reported, sped up, no trial counts. Never present them as measurements.
- Slide titles are at most 40 characters, and the added demo slides keep a sparse layout — both are
asserted by
validate_standalone.py. - Numbers on slides come only from cited evaluations. No cross-paper success-rate leaderboard.
Status
Version 0.13 (2026-09-15): Control 4–16, Data 17–24, Improvement 25–33, closing 34–35.
35 slides, 18 clips, 55-minute suggested timing. HTML SHA-256 9861057cad1b284c1ce17a6c313fbab02ec0662542ee73c8bb26bc1eda6af57c.
Current scope: research/revision_0_13_spec.md. Fixed comparison is .build/revision_0_13_baseline/,
not older Git HEAD. Chapter spine: output/Chapter_Spine.md. Overview diagrams open 4/17/25;
three original hierarchy figures share 5; responsibility shift replaces Hi Robot results on 6; report architecture is 11; original results are 12–14; ENPIRE idea tree is 28.
Results come from research/results_v0_9.json, generated by .build/derive_results_v09.py.
The independent research/results_v0_9_sources.json pins reviewed source identities and is never regenerated by the normal build.
New source authority: research/evidence_v0_11_sources.json; loader .build/evidence_v011.py rejects byte drift.
Original report captures are pinned by research/report_figures_v0_13_sources.json; loader .build/report_figures_v013.py verifies images and query data. Heatmap export expands original DOM overflow only.
RoboDojo public baselines are reweighted published references, not paired reruns; native Score denominators are 48/50.
Table implementation: .build/results_slides_v09.py; per-cell assertions live in the validator.
Claude/RPent pages are removed from the active talk. Historical evidence remains archived. Astra training recipe is undisclosed. Asim budgets differ. RoboDojo prior/interface/horizon co-vary; RoboLab contains retained slots/retries and historical first-five baselines, never a fresh paired benchmark. Real2Sim acceptance is replay-only. ENPIRE physical metrics allow conditional retries; RoboCasa evaluations run each script once, without reset/retry APIs. Do not combine their denominators or Figure 3/7 times. Original curves retain uncertainty; no estimated RoboCasa bar values. Community demos remain qualitative.
Use .build/standalone_delivery_audit.json, .build/standalone_visual_audit.json and
.build/v0_13_extended_audit.json for actual hash-bound status. v0.13 final QA/review passed with 0 unresolved findings; see research/revision_0_13_review.md. Prior v0.9 source-freeze P2 remains resolved.
No measured rehearsal duration, new GitHub release, push, or PPTX/PDF/ZIP regeneration.
[agent_for_robotics] recent context, 2026-09-14 6:23pm GMT+8
No previous sessions found. Slide 16 now uses our independently grounded diagnostic synthesis, approved by the user. The WeChat article itself remains safety-blocked and unread; do not bypass it or attribute this synthesis to it. Await supplied article text/figures for article-specific claims.
Current presentation polish: merged learned tools on 4; Astra begins on 7; capability summary on 16; Takeaways on 34 and Thank You on 35. Native table emphasis is tested per cell; resources and retained slots are descriptive, not an overall winner.
Data now includes Real-to-sim Replay / Data Rollout and the hand CAD case (23). DexGPT original comparison (24) is a failed physical-acceptance example, not a new detailed study. Control timing (15) separates Asim published query seconds from controller Hz and Go1 simulation-time feedback.