Recoleta
GitHub
Source note

SABER: A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation

Authors
Narsimha Menga, Parikshit Sakurikar, Amirreza Rouhi, Satya Sai Reddy, Anirudh Govil, Sri Harsha Chittajallu, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Data ScalingHuman Video RetargetingRetail RoboticsDexterous Manipulation
Open arXivMarkdown source

Summary

SABER is a retail robotics dataset built from real in-store human video and used to adapt VLA policies to grocery-store manipulation. The paper claims that domain-specific action data raises GR00T N1.6 performance on RoboBenchMart retail tasks from 13.4% to 29.3% mean success.

Problem

  • General robot VLA models have weak coverage of retail tasks such as shelf picking, fridge opening, basket loading, floor retrieval, and handling varied packaging.
  • Collecting robot teleoperation data inside active stores is expensive, disruptive, and hard to scale.
  • The gap matters because a policy trained on broad robot data can fail when the store layout, objects, lighting, occlusion, and action sequence differ from its pretraining data.

Approach

  • SABER records about 100 hours of natural activity in real grocery stores using a head-mounted GoPro and a fixed DreamVu ALIA 360° camera.
  • The dataset turns the video into three action streams: 25K LAPA latent-action episodes from egocentric video, 18.6K dexterous hand-pose episodes retargeted to robot joint space, and 1.2K whole-body motion episodes retargeted to a Unitree G1 humanoid.
  • Human annotators review and correct hand and body pose estimates, then retargeting converts human motion into robot-compatible action targets.
  • The authors post-train GR00T N1.6 with a shared-backbone multi-task recipe over the three SABER streams, plus a small amount of robot-native anchor data and task-aligned seed data.

Results

  • SABER contains 44.8K training samples from about 100 hours of real in-store capture.
  • Dataset composition: 25K latent-action sequences, 18.6K hand-pose trajectories, and 1.2K whole-body synchronized motion sequences.
  • SABER-MM post-training reaches 29.3% mean success across 10 RoboBenchMart retail manipulation tasks.
  • The comparison baseline is simulation-only fine-tuning at 13.4% mean success on the same task set.
  • The claimed gain is more than 2.19x over the fine-tuning baseline.
  • A 10K-sample SABER subset is released publicly under CC BY-NC 4.0.

Link