[Feature request] Broader policy & demo coverage: VLA ckpts + more demos

[Everloom-129]

H, thanks for open-sourcing this benchmark! It's a great platform and the multi-robot + asset pipeline is really nice to work with. 🙏 Here Is my usage exp with feature request:

Current policy / demo coverage I see

• Bundled, ready-to-run learned policies: only the Unitree G1 locomotion controllers (loco.onnx / squat.onnx / homie WBC under lw_benchhub/core/mdp/ckpt and data/ckpts/), could there be more?
• Policy clients without bundled weights: GR00T and openpi under policy/ are just server–client wrappers; We need real ckpt to develop and benchmarking.
• RL: skrl PPO trained from scratch (e.g. LiftObj); no pretrained manipulation checkpoints are shipped. It will be great if we can have some simple, open-box demo.

So today there's no turnkey way to (a) run a trained VLA on a task end-to-end for eval, or (b) see a non-G1 / non-manipulation policy in action without VR teleop hardware.

What I'd love to see

1. VLA trained-policy adaptation: guidance/adapters (and ideally a small reference checkpoint or a HF link) for plugging a trained VLA (GR00T-N1.5 / Pi0 / others) into a task end-to-end, with the observation/action mapping conventions documented (the deploy_policy_*.yml mappings are a good start but hard to use without an example that actually runs).
2. More locomotion / loco-manipulation demos like the standalone G1 walking recorder I opened in [feat] Standalone G1 locomotion (loco.onnx) walking demo recorder #43 : e.g. other robots, squat / turn / loco-manip commands, and headless mp4 export, so the bundled low-level policies are discoverable without VR hardware (currently only pandaomron has a keyboard mapping).

I already proposed #43 (a standalone, teleop-free G1 loco.onnx walking demo that renders headless and exports an mp4)
More demos like this will be helpful

Thanks!
Jie Wang