Editable static project page for I3L: Iterative Interactive Imitation Learning for Whole-Body Mobile Manipulation.
Cloned from https://github.com/i-3l/i-3l.github.io. The interactive additions use plain JavaScript and CSS. There is no build step or runtime package installation.
From this directory, run:
python3 scripts/serve.py 8765scripts/serve.py is a static server with HTTP Range support, so videos can seek and keep
their position when the rollout speed changes. Plain python3 -m http.server also works, but
video seeking does not.
Open http://localhost:8765. Edit a file and refresh the browser to see changes. This serves your local copy; publishing changes to GitHub Pages is a separate step.
index.html— the pagestatic/css/index.css— site styles (Bulma + custom)static/css/interactive-figures.css— component explanation and chart stylesstatic/js/figure-config.js— figure titles, component explanations, and activation datastatic/js/interactive-figures.js— inline component explanations and chartsstatic/css/hardware-viewer.css— hardware section and responsive 3D viewerstatic/js/hardware-viewer.js— lazy loading and startup fallbackstatic/js/hardware-scene.js— 3D rendering, joint articulation, and camera controlsstatic/js/hardware-robot.js— matching robot inset for the device explorerstatic/models/— self-contained GLB models exported from the repository CADstatic/images/— figures exported from the paperstatic/videos/walkthrough.mp4— walkthrough video (poster frame:static/images/video_poster.jpg)static/videos/rollouts/— autonomous rollout clips, one per task, each pre-rendered at 1×, 2×, 4×, and 8× (<task>_<speed>x.mp4) from the raw 1× footage with the same crop and trim as the I3L video slides, plus a poster frame (<task>.jpg);static/js/rollouts.jsswaps clips when the speed selector changesstatic/js/takeovers.js,static/css/takeovers.css— "Example Takeover Episodes" player and timelinestatic/js/takeover-data.js,static/videos/takeovers/— generated byscripts/export_takeovers.pystatic/js/takeover-robot.js,static/models/r1pro.glb— 3D R1 Pro beside the takeover video (model fromscripts/export_r1pro.py)static/paper.pdf— drop the compiled paper here (the Paper button links to it)
Every figure follows one pattern: hover, tap, or Tab to part of a figure to highlight it and show its details in a tooltip beside the pointer. Specific numbers live in these tooltips, not in the surrounding text.
- Plots drawn from data — intervention trends, whole-body activation shares, and the operator
study are rendered as SVG from the
chartentries instatic/js/figure-config.js. Hovering a task, operator, or body part highlights it in every panel; legend entries highlight on hover and pin on click. The original PNG stays in the page as the no-JavaScript and print fallback. - Photo figures — the teaser, hardware, tasks, and task-stage figures outline the hovered region with an orange border (the rest of the image is not dimmed), using hotspots from
the
hotspotsarrays. Each entry is["Title", [left, top, width, height], "Short note shown in the tooltip."], with bounds in normalized image coordinates (0–1). - Follow a task — the tasks figure and the task-stage figure outline a task on hover (no
tooltip). Clicking one (or a task in the intervention-trend legend) highlights that task across
the page: both figures, its rollout clip, its table and explorer rows, its trend line, and its
body-part bar. A "Following …" bar at the bottom clears it (or press Escape). These figures use
tasks: [[name, bounds], ...]infigure-config.js; names must match the task names used in the tables and charts. - Tables — the success and strategy tables also render as bar explorers with value tooltips; the published tables stay available under "View published table".
Data sources: trend values were extracted from the vector paths of the paper figure
(intervention_trends_horizontal.pdf) and match the published 63.4% / 69.5% / 72.0% reductions;
operator-study values are computed from figures/operator-study/source-data.json in the paper repo;
activation shares are the printed labels of Fig. 6.
- Takeover timeline: one Round 1 correction episode per task from the
I3L Hugging Face datasets. A badge and a two-row timeline
show policy vs. human takeover (
int_state == 2) at every frame, plus hand-annotated subtasks; click or drag the timeline to seek. To change an episode or a subtask boundary, editTASKSinscripts/export_takeovers.pyand rerun it (needshuggingface_hub,pandas,pyarrow, andffmpeg). It also writes<task>.joints.bin, the Int16 joint track that poses the 3D robot. Seeking needs a server with HTTP Range support (GitHub Pages has it;python3 -m http.serverdoes not). - 3D robot pose:
static/js/takeover-robot.jsrendersstatic/models/r1pro.glbwith the vendored Three.js, only when the section is near the viewport and the screen is wider than 600px, and only re-renders when the video frame changes. The model comes from the OmniGibson soft-gripper USDA viascripts/export_r1pro.py, run inside Blender 4.5 (which welds, decimates, rebuilds the joint tree and, with Node 18+ on PATH, meshopt-compresses with gltfpack). The USDA's JPEG texture atlases must sit in amaterials/folder beside it; the CAD normals are unreliable, so the viewer shades per face. - JoyLo+ episode inset: the lower-right corner of the R1 panel shows both leader
arms, synchronized to the same recorded robot joint track. A blue Policy border
becomes orange with a Takeover label at the annotated intervention frames. This
is a visualization derived from robot-arm poses, not measured leader telemetry
(including for the simulation episode).
takeover-joylo.jsconverts robot angles with the R1 Pro controller's encoder signs and right-arm reflection; the geometric assembly is shared with the hardware explorer injoylo-assembly.js. The inset follows the robot camera and upper-body orientation, fitting closely around the current pose with padding for extended arms. The robot panel is also shown below the video on phones. Both views use the existing renderer/frame clock. If its model fails to load, the robot and video continue without it.
Original images and tables remain readable with JavaScript disabled. Printing shows the original figures and published tables.
Replace images under static/images/, or update their paths in index.html.
Images with teaser-image or method-image receive spotlight regions (or an SVG
chart) from figure-config.js. Customize their titles, hotspot bounds, and notes
there. Bounds are [left, top, width, height], normalized to the original image
dimensions, with all coordinates between 0 and 1.
The success and strategy charts read their values directly from #success-results
and #strategy-results in index.html. Edit those tables to update the charts.
Preserve the group-row class and the table structure. Success cells may contain
percentages or counts with percentages, such as 23/25 (92%); strategy cells
contain numeric percentages. An invalid numeric cell leaves that original table
visible instead of creating a misleading chart.
Body-part activation values in figure-config.js are transcribed from the printed
labels in static/images/whole_body_takeover.png (paper Figure 6). The absent
popcorn torso segment is zero. Published rounding is preserved, including totals
of 99.9%.
Intervention-trend and operator-study values live in the chart entries of
figure-config.js (sources listed under Interactive figures). The task-stage
curves (correction.png) have no source data in the repository, so that figure
uses spotlight regions only; no values have been estimated from pixels.
The optional tests cover hover/focus tooltips and highlighting, figures staying inline, task/series filters, published values, CSV export, mobile layout, printing, and the no-JavaScript fallback. They start their own temporary server and block external requests.
python3 -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
.venv/bin/playwright install chromium
.venv/bin/python -m unittest discover -s tests -vTo use an existing Chrome installation instead of downloading Chromium:
I3L_CHROME=/usr/bin/google-chrome .venv/bin/python -m unittest discover -s tests -vThe Hardware link goes to #hardware. The viewer loads the R1 Pro
leaders as the section approaches the viewport; Franka loads only when selected.
R1 Pro shows a bimanual pair with 14 independent joints. Use Left arm / Right
arm to choose which seven joints to adjust; both arms remain visible and retain
their poses. Franka shows one seven-joint leader.
The upper-right buttons inside the preview switch configurations. A lower-right
robot inset follows the joint sliders, reset, and orbit direction using the same
renderer. R1 Pro uses the inverse of the takeover leader mapping. Franka's CAD is
rooted at its stationary mounting bracket, so its first joint rotates the arm.
franka-joint-map.js maps the CAD assembly offsets and axis directions to Panda
joint coordinates; controls use the intersection of both models' joint limits.
The display home has an upright upper arm and horizontal forearm in both models.
These are model coordinates, not calibrated motor commands. Robot models
load only with their corresponding device; an unavailable inset leaves the device
controls usable. The Panda asset is converted from PyBullet's Apache-2.0 model;
its source, modifications, and license are in static/models/franka-NOTICE.txt
and static/models/franka-LICENSE.txt. Rebuild it with
python scripts/export_franka.py /path/to/pybullet_data/franka_panda --compress
(requires NumPy, trimesh, and Node >= 18 for gltfpack).
Drag to orbit, right-drag to pan, and scroll or pinch to zoom. Camera buttons inside
the lower-left corner of the viewer and
keyboard controls provide alternatives: focus the canvas and use arrow keys to
orbit, +/− to zoom, and 0/Home to fit. The seven labeled sliders move joints within
the source URDF limits. Reset pose restores both arms' illustrated starting poses and
camera. Auto-rotate is opt-in and pauses offscreen, in background tabs, and while
printing. Nothing animates automatically, including with reduced motion enabled.
The original hardware figure remains available if JavaScript or WebGL is unavailable. A failed model load can be retried; a failed configuration switch keeps the current model. Printing uses the original hardware figure.
The assets come from the sibling I3L repository:
assets/gello_r1pro_urdf/robot_7dof.urdfand its referenced STL meshesassets/gello_franka_urdf/robot.urdfand its referenced STL meshes
Exports retain visual colors, link transforms, joint axes, and joint limits.
Positions use 16-bit quantization per mesh and normals use signed 8-bit
quantization; all source triangles are retained. Meshopt compression reduces the
downloads to about 4.6 MB (R1 Pro) and 2.2 MB (Franka). The files use standard
KHR_mesh_quantization and EXT_meshopt_compression glTF extensions. Joint
metadata is stored in node extras.joint. Collision/inertial geometry is omitted;
the viewer is an illustration, not a physics or collision simulator. The R1 Pro
starting pose bends joints 2 and 4. Franka starts in an upright home preset with
a horizontal forearm and downward grip, and Reset pose returns to that preset.
Its CAD-relative joint angles are [0, 64.288, 60.965, 125.495, -25.481, -94.564, -45]
degrees; these include the exported assembly offsets and are not robot motor
commands. The display rotation maps the Franka source model's +X-up axis onto
the viewer's +Y-up axis.
For R1 Pro, the viewer clones the left-leader assembly and reflects it across
the display's sagittal plane to form the right leader, matching the bilateral
setup described in iiil/configs/gello/gello_r1pro.yaml. The two hierarchies share
mesh buffers but have independent joint transforms. The 0.24 m mount spacing
and connecting bar are illustrative, not calibrated assembly dimensions. Model
downloads remain the original individual leader exports.
The BOM also offers separate ZIP downloads of the custom link and grip STL files for each single-arm leader. These use millimetres for slicer imports and include a part list with quantities and dimensions. The archives exclude servo assemblies, hinges, electronics, fasteners and the illustrative bimanual bar. They preserve the source triangles and orientations; print settings and supports must be selected in the slicer. Regenerate these downloads with:
python3 scripts/export_printable_parts.py ../FrankaTrain/I3L/assetsTo regenerate the viewer models from updated CAD, run from this website directory:
python3 scripts/export_hardware.py ../FrankaTrain/I3L/assets
node scripts/compress_hardware.mjsThe first command needs Python 3.9+ (standard library only); the second needs
Node 18+. Neither is needed to serve the website. The exporter records source
and output SHA-256 hashes in static/models/sources.json; compression updates the
output hashes. Both generation scripts and the encoder are checked in.
The runtime uses local, pinned copies of Three.js 0.180.0 and meshoptimizer 0.24.0,
with licenses and provenance under static/js/vendor/. It makes no external
requests for models, decoders, or renderer code. The hardware browser tests load
the real models and check lazy loading, articulation, limits, reset, switching,
downloads, camera controls, mobile touch, printing, retry, and no-JS/no-WebGL
fallbacks. Headless hardware tests enable Chromium's software WebGL renderer.
static/js/analytics.js records page loads and deliberate figure, chart, link,
download and hardware interactions through the site's Cloudflare collector.
Passive observations include section reach, scroll depth, approximate active time,
model load time and viewer errors. The dashboard and database are private.
No cookies or persistent browser identifiers are used. Five reloads count as five
pageviews and an estimated single visitor within the same UTC day. Multi-day
visitor estimates are visitor-days. Location is approximate, and blockers, bots
and network failures affect coverage. Do Not Track, Global Privacy Control and
?analytics=off disable tracking.
Clear the analytics script's data-endpoint in index.html to disable collection.
Keep account credentials, dashboard configuration, and raw data out of this public
repository. Collector failure must never prevent the page or viewer from working.