Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Genie Centurion: Accelerating Scalable Real-World Robot Training with Human Rewind-and-Refine Guidance

Wenhao Wang, Jianheng Song, Chiming Liu, Jiayao Ma, Siyuan Feng, Jingyuan Wang, Yuxin Jiang, Kylin Chen, Sikang Zhan, Yi Wang, Tong Meng, Modi Shi, Xindong He, Guanghui Ren, Yang Yang, Maoqing Yao

TL;DR

GCENT tackles the data bottleneck in real-world robot policy learning by introducing a rewind-and-refine framework guided by a Task Sentinel. The method deploys imperfect policies, triggers human corrections only on failures, and rewinds to collect focused corrective demonstrations, enabling scalable one-to-many supervision. Real-world experiments across four tasks show GCENT achieves up to 40% higher task success and uses less than half the data compared to baselines, with significant data-efficiency gains and reduced operator workload. The approach blends a DAgger-inspired loop with autonomous failure detection to enable scalable, deployable robotics, with potential for large-scale multi-robot supervision and more efficient real-world learning.

Abstract

While Vision-Language-Action (VLA) models show strong generalizability in various tasks, real-world deployment of robotic policy still requires large-scale, high-quality human expert demonstrations. However, data collection via human teleoperation requires continuous operator attention, which is costly, hard to scale. To address this, we propose Genie Centurion (GCENT), a scalable and general data collection paradigm based on human rewind-and-refine guidance, enabling robots' interactive learning in deployment. GCENT starts at an imperfect policy and improves over time. When the robot execution failures occur, GCENT allows robots to revert to a previous state with a rewind mechanism, after which a teleoperator provides corrective demonstrations to refine the policy. This framework supports a one-human-to-many-robots supervision scheme with a Task Sentinel module, which autonomously predicts task success and solicits human intervention when necessary. Empirical results show that GCENT achieves up to 40% higher task success rates than state-of-the-art data collection methods, and reaches comparable performance using less than half the data in long-horizon and precise tasks. We also quantify the data yield-to-effort ratio under multi-robot scenarios, demonstrating GCENT's potential for scalable and cost-efficient robot policy training in real-world environments.

Genie Centurion: Accelerating Scalable Real-World Robot Training with Human Rewind-and-Refine Guidance

TL;DR

GCENT tackles the data bottleneck in real-world robot policy learning by introducing a rewind-and-refine framework guided by a Task Sentinel. The method deploys imperfect policies, triggers human corrections only on failures, and rewinds to collect focused corrective demonstrations, enabling scalable one-to-many supervision. Real-world experiments across four tasks show GCENT achieves up to 40% higher task success and uses less than half the data compared to baselines, with significant data-efficiency gains and reduced operator workload. The approach blends a DAgger-inspired loop with autonomous failure detection to enable scalable, deployable robotics, with potential for large-scale multi-robot supervision and more efficient real-world learning.

Abstract

While Vision-Language-Action (VLA) models show strong generalizability in various tasks, real-world deployment of robotic policy still requires large-scale, high-quality human expert demonstrations. However, data collection via human teleoperation requires continuous operator attention, which is costly, hard to scale. To address this, we propose Genie Centurion (GCENT), a scalable and general data collection paradigm based on human rewind-and-refine guidance, enabling robots' interactive learning in deployment. GCENT starts at an imperfect policy and improves over time. When the robot execution failures occur, GCENT allows robots to revert to a previous state with a rewind mechanism, after which a teleoperator provides corrective demonstrations to refine the policy. This framework supports a one-human-to-many-robots supervision scheme with a Task Sentinel module, which autonomously predicts task success and solicits human intervention when necessary. Empirical results show that GCENT achieves up to 40% higher task success rates than state-of-the-art data collection methods, and reaches comparable performance using less than half the data in long-horizon and precise tasks. We also quantify the data yield-to-effort ratio under multi-robot scenarios, demonstrating GCENT's potential for scalable and cost-efficient robot policy training in real-world environments.

Paper Structure

This paper contains 16 sections, 2 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Learning-Based Manipulation and the Role of Data
  4. Interactive Imitation Learning and Human-in-the-loop Supervision
  5. GCENT Data Collection System
  6. Hardware Setup
  7. Data Processing Pipeline
  8. Task Sentinel: A Multimodal Large Language Model-based Robot Step Detection Model
  9. Rewind and Refine Mechanism
  10. Experiments
  11. Experimental Setup
  12. Q1: Can GCENT Lead to Better Performance?
  13. Q2: Can GCENT Improve Data Efficiency?
  14. Q3: How do rewind strategies make impact?
  15. Fine-tune Details
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Teleoperation Setup illustrating button assignments for the dual 6-DoF controllers and their functions in GCENT.
  • Figure 2: Labels in sandwich assembly task. It shows the mode transitions among four core interaction modes: "Refine", "Rewind", "Inference" and "Awaiting Intervention"(failure).
  • Figure 3: Four tasks in experiments. Two long-horizon and two precise tasks.
  • Figure 4: Comparison of data efficiency across methods. GCENT achieves 0.9+ task score with significantly fewer frames. At the same frame count, GCENT improves model performance by an average of 40%; at the same performance level, GCENT requires only 44.5% of the frames compared to teleop on average.
  • Figure 5: Intervention rates decreased consistently across all tasks as iteration rounds progressed. The typing task, however, exhibited a slight rebound in the final round specifically, as rewind mechanism proportions were intentionally increased only at this terminal stage.