Skip to content

RoboDemo v0.1: Evaluation Release

Latest

Choose a tag to compare

@robodemo-2027 robodemo-2027 released this 20 Nov 10:03

RoboDemo Evaluation Release

  • Five selected checkpoints: Full50 240k, C15 90k, C30 160k, C45 260k and C15-wo 100k.
  • Model architecture, complete evaluation code, 360 task-level results and 3,600 episode logs.
  • Worked human-video policy and learned robot-reference retrieval examples.
  • Shared evaluation image, profile-based asset preparation and an Apptainer launcher for managed notebook containers.
  • Resumable retrieval-feature extraction, scorer weights, annotations and compact reference maps.
  • 113 CPU regression tests per policy variant, verified in fresh-install GitHub Actions and inside the image.
  • Existing validation covers all five checkpoints and two-GPU concurrent evaluation.

Model collection | Quickstart | Validation status

The five model repositories and evaluation image also have a v0.1-evaluation tag. Training code remains planned for the next public release. Eight physical GPUs and Docker Engine GPU startup are not claimed as validated; the published image's tested GPU runtime is Apptainer.