KinRT RoboDojo 10k Candidate Artifacts
This package contains the exact inference artifacts used for the paired
RoboDojo stack_bowls candidate evaluation:
| Directory | Configuration | MoE | Training steps |
|---|---|---|---|
kinrt_10k |
kinrt_lora_robodojo |
4 experts, Top-1 | 10,000 |
pi05_baseline_10k |
pi05_lora_robodojo |
disabled | 10,000 |
Both runs use the same 100 demonstrations, three RGB cameras, 14-D joint
state/action representation, LoRA ranks 32/64, batch size 32, seed 0, and
normalization key RoboDojo-KinRT-stack_bowls100-arx_x5-joint. KinRT alone
uses K=4 action-and-velocity router labels and balanced sampling.
The source revision is
108368422539b79d2be2c596750de1fab1cfd8fd.
The XPolicyLab adapter and complete commands are submitted separately under
policy/KinRT.
Each artifact contains params, assets, checkpoint metadata, and the commit
marker needed for inference. Optimizer and training state are intentionally
excluded. Verify files with sha256sum -c SHA256SUMS after download.
Scope
These are initial 10k submission candidates, not the future 60k full-budget models. Local public-layout results do not replace RoboDojo's official three-seed and hidden-layout verification.
License
KinRT and OpenPI-derived code are released under Apache-2.0. The fine-tuned weights derive from the upstream pi0.5/PaliGemma model family; users must also comply with the applicable upstream model and weight terms. RoboDojo data is not redistributed in this repository.