Published · May 2026

SABER.

A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation — the first high-fidelity retail robotics action dataset built from natural human behavior, not teleoperation.

The Core Claim

Domain-specific robot deployment is fundamentally a data problem. High-fidelity naturalistic human behavior — systematically captured and retargeted — is a scalable foundation for robot adaptation. No robot in the loop required.

44.8K Training Samples

100+ Hours Captured

2.19X Improvement

Resources

[**Watch Videos**
In-store capture demos](https://dreamvu.ai/saber/#videos)
[**arXiv**
Research Paper](https://arxiv.org/abs/2605.09613)
[**Download PDF**
Full research paper](https://arxiv.org/pdf/2605.09613)
[**Dataset**
SABER-10K on Hugging Face](https://huggingface.co/datasets/DreamVu/SABER-10K)
[**Benchmark Results**
RoboBenchMart evaluation](https://dreamvu.ai/saber/#results)

Full Pipeline

Complete SABER Capture Pipeline

Synchronized dual-stream footage: egocentric video, 360° exocentric view, hand landmarks, body skeleton, and SOMA-X mesh — derived simultaneously from real in-store human actions.

The Challenge

Why Retail Demands Its Own Data

Modern VLAs like GR00T N1.6 achieve near-zero success on retail tasks out of the box — not because the model is weak, but because the retail domain is entirely absent from training data.

Distinct Skill Distribution

Articulated object interaction, multi-height shelf reaching, basket loading, floor retrieval, and context-dependent placement — all repeated across hundreds of SKUs in layouts no lab can replicate.

Long-Tail Scene Variation

Dense shelves, active restocking, occlusions, varied lighting, reflective packaging, and product deformability create real-world complexity that generic datasets cannot approximate.

Repetition Matters

A model must see skill families repeatedly across contexts — grasping bottles from different shelf heights, opening fridges from varied approach angles — to achieve reliable deployment.

Performance

Key Results at a Glance

2.19X Improvement over fine-tuning baselines on RoboBenchMart
29.3% Mean success rate across all 10 retail manipulation tasks
91% Average fridge task success — up from 43% baseline
100% Non-robot data — entire dataset captured from human video alone
44.8K Total Samples
100+ Capture Hours
3 Action Streams
10 Eval Tasks

Dataset Architecture

Three Complementary Action Streams

From the same dual-camera in-store captures, three distinct supervision signals are derived — each encoding a different level of kinematic abstraction.

Stream 1

LAPA Latent Actions

25K Embodiment-agnostic motion tokens derived via inverse-dynamics encoding from egocentric video. Captures whole-arm motion, reach trajectories, and grasping dynamics without robot joint labels.

Egocentric GoPro

Stream 2

Dexterous Hand Retargets

18.6K 21-point hand landmarks estimated, human-corrected frame-by-frame, then retargeted to robot joint space via Dex-Retargeting. Provides explicit finger-level precision supervision.

Egocentric GoPro

Stream 3

Whole-Body Retargets

1.2K SOMA-X body parameters estimated from the 360° ALIA view, human-corrected, and retargeted to the Unitree G1 humanoid. Provides torso-arm-leg coordination for floor retrieval and extended reach.

Exocentric ALIA 360°

Methodology

From Store Footage to Robot Training

SABER is constructed from a dual-stream capture architecture — egocentric GoPro + exocentric ALIA 360° — across multiple real grocery stores.

1. In-Store Capture

100+ hours across multiple real grocery stores with head-mounted GoPro + DreamVu ALIA 360°

2. Action Extraction

LAPA encoding, hand pose estimation, and SOMA-X body estimation with human QC annotation

3. Robot Retargeting

Dex-Retargeting to robot hand joint space + SOMA-X-to-Unitree G1 whole-body retargeting

4. VLA Post-Training

Shared-backbone multi-task training on GR00T N1.6 with flow-matching objective

Demo Videos

Capture Sessions & Task Annotations

Annotated in-store capture footage from the SABER dataset — showing retail manipulation tasks with action labels and multi-scene diversity.

Retail Task Cycles

Pushing trolleys, packing goods, arranging goods, opening doors, inspecting labels, and handling baskets.

Retail Task Cycles

Placing and moving foods, scooping loose goods, inspecting deformable packets, carrying multiple goods, inspecting fruits, closing doors, and placing goods.

Evaluation

RoboBenchMart Results

SABER-MM post-training on GR00T N1.6 evaluated across 10 retail manipulation tasks spanning fridge, board-to-board, floor pick, and basket pick categories.

Mean Success — All Tasks

29.3%
13.4% Fridge Tasks (avg open + close)
91%
43% Floor Pick Tasks (avg)
17%
3%
2.19X Mean improvement over baseline

Training Corpus

SABER-MM Data Composition

The post-training corpus combines SABER's three streams with robot-native anchor data and task-aligned demonstrations — totaling ~52.1K samples.

52.1K Total Samples
SABER — LAPA Latent Actions
25K samples · Egocentric video
48.0%
SABER — Hand Retargets
18.6K samples · Dex-Retargeting
35.7%
SABER — Body Retargets
1.2K samples · Unitree G1
2.3%
NVIDIA Robot Data
4.8K samples · Anchor signal
9.2%
RoboBenchMart
2.5K samples · Task-aligned
4.8%

Key Insights

What SABER Demonstrates

Finding 01

SABER demonstrates that high-fidelity naturalistic human behavior, systematically captured and retargeted, is a viable and scalable foundation for domain-specific robot adaptation — without a robot in the loop.

Finding 02

Three Streams Are Complementary

LAPA tokens capture whole-arm trajectory, Dex-Retargeting provides finger-level precision, and body retargets supply torso-arm-leg coordination. Together they provide non-overlapping kinematic information.

Finding 03

Robot-Native Anchor Stabilizes Training

The 4,800-sample robot-native anchor data proved necessary to stabilize early training even at SABER's scale, suggesting general manipulation signal matters for robust convergence.

Finding 04

Task Progress Beyond Binary Success

SABER-MM teaches models to progress further through each task sequence — mean P≥2/3 of 0.445 vs 0.278 baseline — indicating reaching and grasping are well-learned while placement remains the frontier.

Citation

Cite This Work

Copy BibTeX

@article{dreamvu2026saber,
  title   = {SABER: A Scalable Action-Based Embodied Dataset
             for Real-World VLA Adaptation},
  author  = {Menga, Narsimha and Sakurikar, Parikshit and Rouhi, Amirreza
             and Reddy, Satya Sai and Govil, Anirudh and Chittajallu, Sri Harsha
             and Aggarwal, Rajat and Namboodiri, Anoop and Reddi, Sashi},
  year    = {2026},
  month   = {May},
  note    = {DreamVu Inc.},
  url     = {https://dreamvu.ai/saber}
}