- CSS 52%
- JavaScript 41%
- Vue 6%
- PowerShell <1%
- Shell <1%
Towards Adaptable Partners — HIRO Lab Slidev Talk #
A native, editable Slidev repository that automatically downloads the original HIRO Lab theme.
Talk: Towards Adaptable Partners on the Road, in the Factory and in the Home
Presenter: Naren Dasan, Human Interaction and RObotics (HIRO) Group, CU Boulder
The theme is cloned from https://github.com/narendasan/hiro-lab-slidev-theme at first launch. It is not a PowerPoint recreation or a substitute CSS theme. Slidev loads the actual upstream source from vendor/hiro-lab-slidev-theme.
Quick start #
Install Node.js 20+ (includes npm) and Git once on your computer. An internet connection is required for the initial install to download the theme from GitHub and Slidev dependencies from npm. No global Slidev installation is necessary.
macOS / Linux
unzip HIRO_Adaptive_Autonomy_Slidev_Repo.zip
cd HIRO_Adaptive_Autonomy_Slidev_Repo
./start.sh
Windows PowerShell
Extract the ZIP, enter its folder, then run:
.\start.ps1
Alternatively, on any platform with Node.js and Git available:
npm start
The first command automatically:
- Clones
narendasan/hiro-lab-slidev-theme(branchmaster) intovendor/hiro-lab-slidev-themeusing Git. - Verifies the cloned package is the expected
slidev-theme-hirotheme and that all referenced paper images are present. - Installs the project's Slidev packages using npm (the theme is used directly from the local clone).
- Opens the editable Slidev presentation in your browser.
Subsequent launches reuse the downloaded theme and installed packages. Installation requires Node and Git but does not require manually installing or cloning a separate theme repository. The final URL is printed by Slidev (typically http://localhost:3030).
Common commands #
| Command | Result |
|---|---|
npm start or ./start.sh |
Automatically install if needed, then run slides. |
npm run setup or ./setup.sh |
Download theme and install packages without launching slides. |
npm run dev |
Start quickly after setup; no bootstrap checks. |
npm run theme:update |
Pull newest theme changes from upstream and reinstall if necessary. |
npm run build |
Produce static HTML in dist/ after setup. |
node scripts/bootstrap.mjs --build |
Install if needed, then build the static site. |
npm run export |
Export a portable PDF to output/pdf/ using Slidev and Playwright Chromium. |
npm run check |
Verify theme presence and local assets after setup. |
npm run test:navigation |
Check that the slide-search list stays empty when closed. |
npm run test:setup |
Offline smoke test of theme cloning and launch automation (POSIX). |
Project layout #
HIRO_Adaptive_Autonomy_Slidev_Repo/
├── slides.md # Editable talk; custom Slidev layouts and notes
├── style.css # Talk-specific CSS additions
├── public/images/ # Figures extracted from the supplied papers
├── public/brand/ # Original theme logos and license
├── package.json # Slidev dependencies and commands
├── scripts/
│ ├── bootstrap.mjs # One-command theme install / start / build
│ ├── check.mjs # Validate cloned theme and all image paths
│ └── test-bootstrap.mjs # Offline installer smoke test
├── vendor/
│ └── README.md # Theme is cloned here on first launch
├── start.sh, setup.sh # macOS / Linux convenience launchers
├── start.ps1, setup.ps1 # Windows convenience launchers
└── .github/workflows/verify.yml # CI builds with the real theme
vendor/hiro-lab-slidev-theme/, node_modules/ and dist/ are gitignored, since they can be restored by the installer. This keeps the project small and avoids silently bundling a modified copy of the theme.
Edit the talk #
Edit slides.md; Slidev hot-reloads changes automatically. Keep per-slide layout: and class: directly after ---, without an intervening blank line. The deck is excluded from ordinary Prettier formatting because it can split frontmatter into extra slides. Each --- separator begins a new slide. Paper figure files live in public/images. Styling tweaks in style.css extend, but do not replace, the upstream theme. Presenter notes are included as comments in the Markdown.
The deck begins with unpublished NVIDIA work on online driving-intent priors, followed by preference transfer (IEEE RA-L 2024) and ABR² online policy repair (ICLR 2027, under review). The road section begins with a merge scenario, then introduces NVIDIA Safety Force Field in one slide as a prior-free safety baseline: no driver-specific intent prior is required, while driving quality needs additional criteria. A second merge slide follows SFF, showing how comfort, timing, and an intent prior shape a smaller workable maneuver set within the same hard safety boundary. Per-actor online prior estimation and a formal subset relationship follow. Actor-specific histories and beliefs are maintained only while relevant to the interaction, then evicted when actors exit or otherwise cease to be relevant. Alpamayo follows as a possible modern route to in-context intent estimation at fixed model weights, immediately before its “Thinking Out Loud” demo. The presenter supplies the chronology that this reasoning-VLA route was unavailable during their NVIDIA work; the talk does not claim a validated Alpamayo actor-prior estimator. Zero-shot here means responding to an unfamiliar actor before learning its driving style. Safety remains conditional on the model assumptions; efficient coordination and passenger comfort are separate objectives for online adaptation. The SFF overview follows NVIDIA's white paper and DRIVE Labs article; the safety-set notation summarizes its admissible-control role. The unpublished work concerns a belief-conditioned interaction workspace, while the hard SFF boundary remains distinct. Presenter notes connect the driving setting to two motivating gaps: learning about a partner from limited evidence, and improving the agent's own behavior when prior competence is insufficient. These motivate the later preference-transfer and policy-repair sections as conceptual research connections, without asserting unpublished AV performance results or a historical order of project origins. The main talk ends at the synthesis slide, followed by references and three technical backups. The ABR² learning-curves slide includes a final-evaluation example from Table 2 before the scene-perturbation limitation. Every research slide includes a source and presenter notes.
The three local rollout videos appear together immediately after the ABR² training explanation. Click each player’s controls to play or pause. Videos pause when leaving the slide and reset when revisited. The files and their poster images live in public/videos/ and are included in static builds. PDF exports show still previews; play the videos from Slidev.
npm run export renders each slide to a high-resolution image before assembling the PDF. This preserves the theme’s gradient text across PDF viewers. The PDF is for presenting and review; edit text, equations and tables in slides.md.
The original HIRO cover, dark cover, section, two-column and footer layouts are reused. Talk-specific compositions and fixed type sizes live in style.css, without edits to the downloaded theme. The bundled Geom font keeps the deck readable offline. Original theme logos are copied to public/brand/ so both development and static builds resolve them.
Paper figures are high-resolution excerpts of the supplied PDFs. Preference-study percentages describe human-action prediction accuracy; ABR² percentages describe simulated robot task success. The scene-perturbation limitation remains in the main talk, with the complete table in backup.
Theme and updates #
The theme's package.json identifies it as slidev-theme-hiro. The slides.md frontmatter points to the local folder as recommended by Slidev:
theme: ./vendor/hiro-lab-slidev-theme
Theme updates are opt-in (npm run theme:update), so a previously working talk will not unexpectedly change appearance between presentations. By default, master is cloned with shallow Git history. For a mirror or a local test checkout, set HIRO_THEME_URL when running npm start; optionally set HIRO_THEME_BRANCH (defaults to master).
If the theme clone is incomplete, remove vendor/hiro-lab-slidev-theme and rerun npm start. If a direct npm run dev command reports a missing theme, use npm start to bootstrap first. For presentation rooms with limited connectivity, run setup and npm run build before arriving.
Source material and licensing #
- Nemlekar, H., Sivagnanadasan, N., et al. (2024). Selecting Source Tasks for Transfer Learning of Human Preferences. IEEE Robotics and Automation Letters, 9(8): 6896–6903. DOI: 10.1109/LRA.2024.3415432.
- Nemlekar, H., Dhanaraj, N., Guan, A., Gupta, S. K., and Nikolaidis, S. (2023). Transfer Learning of Human Preferences for Proactive Robot Assistance in Assembly Tasks. HRI 2023, 575–583. DOI: 10.1145/3568162.3576965. The last slide of section 2 covers online refinement of the transferred model.
- Naren Sivagnanadasan, Andrew Silva, Joshua Haskins, Rin Metcalf, Barry-John Theobald, and Alessandro Roncone. ABR²: Advantage-Based Regularization for Online Policy Repair. Under review at ICLR 2027; source manuscript provided for this talk. Sivagnanadasan, Haskins, and Roncone: Department of Computer Science, University of Colorado Boulder. Silva, Metcalf, and Theobald: Apple Inc. Author attribution and affiliations supplied by the presenter.
The theme is the author's independent NCSA-licensed project, fetched from GitHub. The original logo assets in public/brand/ carry the upstream license in that folder. Figure reproductions are sourced from the listed papers and should retain academic attributions if the slides are distributed. See THIRD_PARTY_NOTICES.md.
Build verification #
npm run test:setup tests bootstrap logic with a local fake theme and a mock npm installer, so it can run without GitHub or npm registry access. Full live theme build verification additionally requires network access; the provided GitHub Actions workflow performs npm run setup and npm run build when the repository is pushed to GitHub.
Slide navigation compatibility #
The fuse.js dependency used by Slidev 0.49 is pinned to 7.1.0 through an npm override. Later releases return all items for an empty search, which exposes Slidev’s offscreen “Go to slide” suggestion list over the presentation even when closed. The pin preserves normal title/number search. After dependency changes, restart the development server and refresh the browser. npm run test:navigation checks this behavior.
Alpamayo demonstrations #
The NVIDIA section includes the official “Thinking Out Loud” YouTube demo linked from NVIDIA’s Alpamayo product page. Click the local preview to load the player. The demo requires internet, pauses when leaving the slide, and shows a local still in the PDF. The three ABR² rollout clips remain local.
The lane-merge illustration is an AI-generated conceptual scene, not NVIDIA data. The current slide uses public/images/lane_merge_futures.png, with dotted paths showing alternative futures; its edit prompt and provenance are recorded in the adjacent .prompt.txt file. The original scene is retained as lane_merge_illustration.png. The second version, lane_merge_workspaces.png, compares schematic workable regions under claiming and yielding priors; its prompt and provenance are stored alongside it.