Spaces:

---

lerobot

/

steer-a-robot

Like 30

Running on Zero

Redesigns the Space's interface in the LeRobot brand and fixes session-control bugs. Robot physics, policy inference, the browser renderer, and the exported scene assets are unchanged.

Screenshots

Local preview: static scene, simulated session states.

Interface

• New layout, moved from app.py into ui.py and shared with the local preview.
• LeRobot look: the official logo (README banner, scaled, tracked with Git LFS), the logo's orange #FF9D00 and yellow #FFD21E, black outlines with hard shadows, and Hugging Face's typefaces (Source Sans 3, IBM Plex Mono). The hero gets the full treatment; the workspace stays calm, with yellow marking what is active. Light mode is forced.
• Phone layout uses the full width. Gradio prefixes custom selectors, which had silently disabled the phone padding rule and left 236 px of content on a 390 px screen.
• A slim status bar above the simulated view shows the state, the current instruction, the goal or note, the timer, and an inline Stop button. It keeps a fixed height, so the view never shifts.
• The skill library highlights the active trained instruction.
• Task reasoning and Scene Q&A are separate conversations.
• Example goals fill the goal box; Steer sends them.
• The top and bottom LeRobot links share one button style. Gradio's own footer link uses the class .built-with too, so it had been restyled as the prominent button.
• Gradio's per-output progress lines are hidden for the streaming session event; the status bar already shows progress.

Behavior fixes

• Questions never start a session. LeRobot's RolloutController.ask() only answers during a run segment, so a question used to start the robot on the scene's default instruction (task 8, "put the bowl on the plate"). The Ask box is now disabled between sessions, and send() refuses a question that arrives without one.
• No more blanked controls. For a command that doesn't start a session, run() now yields one gr.skip(). A generator that yields nothing makes Gradio set every output to None, which blanked the status bar and the skill-button labels after an empty submit, and briefly after every follow-up command during a session.
• If a session fails, the controls reset to an error state.

API changes

• The example_0–example_3 endpoints are removed; examples now run in the browser only. steer, ask, stop, and task_* are unchanged.
• The session event has one more output (the question box).

Development

• scripts/preview.py runs the production UI on a laptop, with mock sessions and no models or GPU.
• make setup, make check, make preview. make check runs compilation, Ruff (Pyflakes rules), a JS syntax check, the GLB/camera asset contract, 16 unit tests, and a dependency check.
• DEVELOPMENT.md covers local setup, full-backend requirements, and a code walkthrough.

Testing

• make check passes (16 tests).
• Exercised in the local preview at desktop and phone widths: status bar states, Stop, example goals, Ask gating, typing while session updates stream, and empty submits.
• Not yet run on the full backend (Linux, CUDA, planner). The app.py changes are small (the send() question guard, the run() skip, the extra output, and the light-theme hook), but this needs a GPU run before merging. After pushing, check that the fonts and logo load on the Space itself.

🤖 Generated with Claude Code