Memory Anchors: dropping top 10% rehearsal data worsens robot forgetting up to 4.5x
SongShuran · x · 2026-11-10
Key finding
A CoRL 2026 paper on continual learning for robots identifies a special subset of rehearsal data — Memory Anchors — that largely determines whether a robot retains or forgets old behaviors when learning new ones.
Results
- Withholding just the top 10% of Memory Anchors from rehearsal increases forgetting by up to 4.5x;
- Increasing their presence reduces task forgetting and enabled a real robot to learn challenging task sequences;
- The work explains why rehearsal works so well for continual learning (especially VLAs) and when it fails.
Poster presented Monday at JW Marriott; paper and website are public.
More from Embodied
- New-gen UMI gripper passes real-world bar test, lifting a full glass of water — YuXiang_IRVL · 2026-11-10
- Unitree G1 humanoid learns to walk from scratch in under an hour on an old GPU — MarcJSchmidt · 2026-11-10
- Steve Waslander appointed Director of U of T Robotics Institute — VectorInst · 2026-11-10
- Boston Dynamics' Atlas robot learns to roller skate and pull off a 360 aerial spin — adariostrange · 2026-11-10
- Embodied Turing Machines: coding agents write reusable robot routines, 70.24% on 42 bimanual tasks — jiqizhixin · 2026-11-10
- CoRL 2026: EgoScale lands oral session, latest GR00T2 results at H2R workshop — chris_j_paxton · 2026-11-10