H2RB native baseline: DexWild / object
Public inference checkpoint for the native DexWild baseline on the H2RB object transfer data split. Fixed final optimizer step: 70000; seed: 42.
Upstream public training implementation: https://github.com/dexwild/dexwild-training at commit efc1643865f2c9265d2351a1ecfd0731368930e3. This is an H2RB adaptation trained by us, not an official checkpoint released by the upstream authors.
TRAINING_FINISHED.json and ARCHIVE_READY.json contain the frozen training and checksum contract. Training completion is not a robot rollout success claim.