Simulated manipulation — trained policy adapters and probes
Model weights trained in simulation for robot manipulation and 2-D navigation experiments.
All weights here were trained by the uploader. Pretrained backbones are not redistributed; the training code downloads them from their own sources under their own licences.
Contents
| Path | What it is |
|---|---|
star/pi0_anchor_lora.pt |
LoRA + projector adapter, two-cube tabletop task |
star/pi0_noplan_lora.pt |
the same adapter trained with the conditioning dropped |
star/pi0_richplan_lora_rp.pt |
adapter trained on a real-prompt demonstration mixture |
star/pi0_richplan_lora_dag.pt |
the above, refined with a privileged-expert DAgger stage |
star/dino_probe_spoon448.pt |
spatial-softmax probe over frozen visual features |
star/p2_unified_ckpt.pt |
a single generator trained jointly over four task families |
star/pi0_general_reasoner.pt |
joint generator over four demonstration sets |
star/robocasa_steer_lora.pt |
adapter for a kitchen-simulator pick-and-place task |
star/adapters_{cf,v2,mixture}/ |
adapters for a bowl-selection benchmark, several data mixtures |
star/heads_cf_fixed/ |
small conditional generators, multiple seeds |
star/dual_channel_head_s0.pt |
a two-channel conditioning head |
maze/saved_models/ |
18 multi-path maze navigation runs, several architectures |
Verifying a download
A manifest pinning every file by content hash accompanies the training code.
Caveat
The exact training invocation for pi0_richplan_lora_rp.pt was not preserved. Use these
weights rather than trying to reconstruct it.