About Me

  • I am Zhang Jinquan(张锦权), an Artificial Intelligence master's student at Shenzhen University and a member of the World Model Group at the Guangdong Laboratory of Artificial Intelligence and Digital Economy (Shenzhen), led by Tian Qi (田奇), Chief Scientist of Huawei Terminal BG. I received my bachelor's degree from Jiangsu University.
  • My research focuses on robot learning and real-world robustness of VLA / WAM models, especially post-training and fine-tuning for physical environments.
  • My work has been accepted by IROS and AHSWN, and I have filed multiple invention patents.
  • I am currently interning at Roboscience, working on VLA / WAM fine-tuning and post-training for strong simulation performance and physical-world robustness, including robot evaluations in large shopping-mall and supermarket shelf scenarios.

I hope to join a dynamic and goal-driven robotics team. If you are interested, please email me at 2410815010@mails.szu.edu.cn.

News

  • 2026.06: My paper Structure-Aware Robust Fine-Tuning: Defending Vision-Language-Action Robots Against Physical Attention Hijacking was accepted by IROS 2026.
  • 2026.05: My paper VLAGuard: A Framework for Evaluating and Mitigating Physical Attention Hijacking in Vision-Language-Action Robots within Wireless Sensor Networks was accepted by the journal AHSWN.
  • 2026.05: My invention patent 基于语义嵌入和注意力双一致性的机器人控制方法、系统、终端及存储介质 was granted. Patent No.: ZL 2026 1 0211734.7.
  • 2026.05: My invention patent 一种基于视觉语言动作模型的机器人控制方法、系统、终端及存储介质 was granted. Patent No.: ZL 2026 1 0219615.6.
  • 2026.05: My invention patent 一种基于视觉语言模型的机器人操作方法、系统及终端 was granted. Patent No.: ZL 2026 1 0200293.0.

Achievements

Papers

  • Structure-Aware Robust Fine-Tuning: Defending Vision-Language-Action Robots Against Physical Attention Hijacking. Accepted by IROS 2026.
  • VLAGuard: A Framework for Evaluating and Mitigating Physical Attention Hijacking in Vision-Language-Action Robots within Wireless Sensor Networks. Accepted by AHSWN.

Patents

  • 基于语义嵌入和注意力双一致性的机器人控制方法、系统、终端及存储介质. Granted. Patent No.: ZL 2026 1 0211734.7.
  • 一种基于视觉语言动作模型的机器人控制方法、系统、终端及存储介质. Granted. Patent No.: ZL 2026 1 0219615.6.
  • 一种基于视觉语言模型的机器人操作方法、系统及终端. Granted. Patent No.: ZL 2026 1 0200293.0.
  • 一种约束驱动轨迹生成方法、系统、终端及存储介质. Status: under substantive examination.
  • 一种柔性压力传感器及其制备方法. Status: under substantive examination.
  • 一种基于未来潜在表征一致性约束的视觉语言动作模型鲁棒训练方法. Status: under preliminary examination.
Patent certificate for ZL 2026 1 0211734.7
Certificate - ZL 2026 1 0211734.7
Patent certificate for ZL 2026 1 0219615.6
Certificate - ZL 2026 1 0219615.6
Patent certificate for ZL 2026 1 0200293.0
Certificate - ZL 2026 1 0200293.0
2 / 3

Selected Publications

Robot Learning and VLA Robustness

IROS 2026
SARF paper overview figure

Structure-Aware Robust Fine-Tuning: Defending Vision-Language-Action Robots Against Physical Attention Hijacking

Zhang Jinquan(张锦权) et al.

  • Identifies policy-critical action-to-vision attention hijacking, where physical adversarial patches divert action-conditioned attention away from task-relevant regions.
  • Proposes AGSD, an EOT-optimized printable patch that jointly hijacks policy-critical attention and disrupts vision-language semantic alignment.
  • Introduces SARF, a zero-inference-overhead defense that fine-tunes only the visual encoder using feature anchoring, attention correction, and language-guided geometric consistency.
  • Reduces OpenVLA failure under AGSD from 100% to 17.0-56.8% on LIBERO and improves real PiPER success from 9.7% to 61.7%.
AHSWN 2026
VLAGuard framework overview figure

VLAGuard: A Framework for Evaluating and Mitigating Physical Attention Hijacking in Vision-Language-Action Robots within Wireless Sensor Networks

Zhang Jinquan(张锦权) et al.

  • Studies VLA robots as mobile edge nodes in WSN-assisted smart environments and evaluates their robustness against physical adversarial threats.
  • Introduces VASA, a printable-patch stress test that distracts action-conditioned cross-attention through policy-critical attention hijacking.
  • Proposes APFT, a zero-inference-overhead defense that stabilizes spatial-temporal attention and enforces geometric consistency.
  • Reduces OpenVLA failure from 100.0% to 25.9% in LIBERO and improves real-world success from 23.0% to 67.4% across 2,000 severe patch-attack trials.

Selected Projects

Real-Robot Learning Systems

Project 2026
Aloha Pi0.5 LeRobot project overview

Aloha Pi0.5 LeRobot: A Real-Robot Learning Pipeline

Zhang Jinquan(张锦权)

  • Builds an end-to-end pipeline for AlohaMini teleoperation, LeRobot-format datasets, dataset cleaning, Pi0.5 training, and 4090-to-Orin remote inference.
  • Separates real-time robot IO and safety on Orin from heavy VLA policy inference on a remote GPU server.
  • Provides public engineering documentation for reproducible real-robot learning without exposing private datasets, paths, IPs, or hardware identifiers.

Research Interests

  • Robot Learning: data-driven manipulation, real-robot systems, and closed-loop deployment.
  • VLA: Vision-Language-Action policies for general robotic manipulation.
  • WAM: World Action Models for action-aware embodied prediction and planning.
  • Post-Training: adapting pretrained robot policies for downstream tasks and robustness.
  • Fine-Tuning: structure-aware and parameter-efficient adaptation for embodied models.

Experiences

Demo Videos

FingerEye-Based Dexterous Manipulation Demo

  • This work builds on Roboscience's FingerEye model, with our own algorithmic improvements for real-robot deployment.
  • During post-training, we found that mixing successful model-inference data with human teleoperation data works better than training only on manually collected teleoperation data.
  • Based on this finding, we designed an automated architecture for collecting successful datasets from the model inference stage.
  • We combine human scoring with scoring from the Qwen vision-language model, producing thousands of high-quality real-robot post-training samples.
  • Compared with training on pure human teleoperation data, the mixed-data setting improved real-robot inference success by nearly 30%, validating our hypothesis.
  • We are accelerating the write-up of this work and hope to share it on arXiv and at ICRA 2027 for discussion on stable and robust real-robot deployment.

FingerEye-Based Envelope Opening Demo

  • This demo extends our FingerEye-based real-robot manipulation work to an envelope-opening task.
  • The robot must coordinate close-range perception and finger-level contact control to grasp, stabilize, and open a flexible envelope.
  • Compared with rigid-object manipulation, this task stresses fine-grained dexterity, deformable-object interaction, and robust execution under small contact changes.
  • We use this task as a practical real-world stress test for the FingerEye project, pushing the policy toward useful everyday manipulation skills.

Unified Pi0.5 LoRA Demo on the Songling PIPER Manipulator

  • This demo shows the same Pi0.5 policy after real-robot LoRA fine-tuning on the Songling PIPER manipulator.
  • All three tasks are completed by a single checkpoint, highlighting cross-task consistency instead of task-specific model switching.
  • We smooth both model outputs and arm execution, because real-world robots should be measured not only by task success, but also by the fluidity and stability of their motions.
  • Our broader goal is to move beyond lab-only demonstrations and build robots that can operate gracefully in everyday human environments.

Qwen + Flow-Matching VLA Generalization Demo

  • This demo explores a VLA policy that combines a Qwen vision-language backbone with a flow-matching action head, connecting high-level semantic understanding with continuous robot control.
  • We first validated the framework in LIBERO, achieving over 90% success across all four suites and reaching 97.6% on the Spatial suite.
  • We further evaluated the same direction on real ALOHA hardware and compared it with the native JAX-format Pi0.5 model under extensive real-robot trials.
  • Across same-category objects with unseen colors, novel distractors entering the robot camera view, and substantially different backgrounds, the Qwen + flow-matching VLA showed stronger cross-object and cross-background generalization.
  • These results suggest that stronger semantic grounding and generative action modeling can improve the path from simulation benchmarks to robust real-world manipulation.

SARF: Robust Fine-Tuning Against Physical Attention Hijacking

  • This demo presents our IROS 2026 work, Structure-Aware Robust Fine-Tuning, for defending VLA robots against physical attention hijacking.
  • We identify a mechanism-level vulnerability where printable adversarial patches divert policy-critical action-to-vision attention away from task-relevant regions.
  • To stress-test this failure mode, AGSD jointly concentrates action-conditioned attention on the patch and disrupts vision-language semantic alignment under realistic transformations.
  • SARF fine-tunes only the visual encoder with feature anchoring, policy-critical attention correction, and language-guided geometric consistency, adding zero inference overhead at deployment.
  • On LIBERO under adaptive AGSD, SARF reduces OpenVLA failure rates from 100.0% to 17.0 / 14.2 / 26.5 / 56.8% on Spatial/Object/Goal/Long, averaging 28.6% failure rate with a 71.4-point absolute reduction.
  • On the real PiPER platform, AGSD drops average success from 72.0% clean performance to 9.7%; SARF restores it to 61.7%, with task success rates of 74.0 / 63.0 / 58.0% for pick-and-place, open-drawer, and stack.
1 / 5

Contact