Release plan for Stage 2 (DAgger/distillation), Stage 3 (HL tennis policy), datasets, and checkpoints

[purenightmare]

Hi LATENT team,

Big fan of the work — we’ve been playing with the tracking pipeline and poking around hl_proxy for Stage 2 (Pulse distill) and Stage 3 (HL PPO). Really impressive stuff.

I checked the README TODO list (DAgger, latent model, HL training code, full mocap, checkpoints, etc.), but didn’t see any expected timeline for those releases. Totally understand if things aren’t fixed yet — just wondering:

1. Is there a rough roadmap (e.g., what’s likely next vs. later)?
2. Do you plan to open-source high-level policy, or should we expect something different?
3. Any plan to release pretrained Stage-2/3 weights and full ball/mocap data alongside the code?
   No rush — even a high-level “coming months / TBD” would help us plan our own experiments. Thanks for open-sourcing what you already have!

[zzk273]

I'm afraid we won't be able to open-source any further in the near term, as company policy does not allow it. For the DAgger part, you may refer to the OpenTrack repository. Sorry for that.