Shipped, active, planned, and future stages stay separate from ambition.
From one public sample to multi-episode embodied-AI experiments.
The roadmap keeps stage timing, research tracks, linked tasks, scale-up gates, and source mirrors visible without a separate track selector.
Human motion, 3D/4D reconstruction, egocentric interaction, and world modeling are research views over the shared task suite.
JSON mirrors and GitHub notes support validation, reproduction, and citation checks.
Loading roadmap artifacts...
01 / Stage timeline
What has shipped, what is active, and what remains future work.
The timeline keeps implementation state separate from ambition: public-sample task work, selected-episode baselines, Qwen3-Omni diagnostics, branch experiments, and the long-range pretraining goal.
02 / Research tracks
Four tracks with their task evidence kept together.
Each track lists the linked 20-task-suite heads, current evidence state, next steps, and extension probes in one card. Nothing is hidden behind a selected-track side panel.
03 / Scale-up gates
Data preparation and model-branch decisions.
This section keeps the selected-episode gate, Qwen3-Omni pilot contract, and foundation-model branch matrix separate from the task cards so the training route stays clear.
04 / 20-task coverage
All task contracts in one index.
The compact task index is for scanning. It shows each task’s output, metric, and direction roles without forcing another navigation layer.
Structured mirrors for scripts and validators are attached here. The sections above summarize the roadmap before these records.