Browser-based visualization platform with per-demo dashboards

Motivation

The browser should be a one-stop shop for everything Geodude — 3D view, chat, sensors, planning diagnostics, and per-demo panels.

Phase 1: Basic integration — DONE

• mj_viser in workspace (personalrobotics/mj_viser)
• --viser CLI flag
• Real-time trajectory sync during execution
• Textures, visibility toggles, hidden objects, STL double-sided rendering

Phase 2: Core panels — replace terminal switching

Chat panel

• Chat input + response display in browser
• Tool call visualization (show pickup/place as they execute)
• Action log (scrollable history)

F/T sensor panel

• Live time-series plot (both arms)
• Toggle: local/world frame, raw/tared
• Tare button + baseline indicator

Scene state panel

• Object list with names, types, positions, active/hidden/held status
• Click object in list to highlight in 3D view

Planning diagnostics panel

• Last plan result: success/fail + reason
• IK unreachable vs in collision breakdown

Phase 3: Interactive features

• Click-to-select objects in 3D view
• Live BT visualization with node status
• Arm state overlay (joint positions, EE pose axes)
• Token/cost ticker

Phase 4: Per-demo dashboards

Demo authors define custom panels in Python. Recycling: bin fill counter, success rate. Stacking: height tracking, alignment error. Force control: compliance viz, force profile.

Phase 5: Advanced

• Drag-to-place (drag object then robot plans pick-and-place)
• Trajectory preview (ghost arm before executing)
• Recording/playback with annotated timeline
• Multi-view (gripper cam, overhead)
• Shareable sessions (send URL, multi-client via Viser)

Start by reviewing the existing mj_viser workspace integration and the --viser CLI flag. The issue describes a multi-phase roadmap spanning chat, sensors, scene state, planning, interactive controls, and custom dashboards, but names no files, tests, or bounded first milestone; a contributor would need the scope narrowed before implementation can be considered done.