🦚 PEAfowl

Perception-Enhanced Multi-View Vision-Language-Action for Bimanual Manipulation

Qingyu Fan1, Zhaoxiang Li2,*, Jinrui Hu1,*, Yi Lu2, Wang Chen2, Qiu Shen2, Xiao-xiao Long2,†, Yinghao Cai1,†, Tao Lu1, Shuo Wang1, Xun Cao2
1Institute of Automation, Chinese Academy of Sciences
2Nanjing University
*Equal contribution   †Corresponding authors
Accepted by IEEE Robotics and Automation Letters (RA-L), 2026

🏆 Award: Grand Prize (3rd Place), AI+ Special Competition (Challenge Track), the 19th "Challenge Cup" (China Mobile), based on an industry problem proposed by Horizon Robotics.

IEEE Xplore DOI arXiv Video Code Deployment Citation
Motivation and overview of PEAfowl

Motivation and overview of PEAfowl.

Overview

PEAfowl is a perception-enhanced multi-view vision-language-action policy for bimanual manipulation. It predicts per-token depth distributions, performs differentiable 3D lifting, and aggregates local cross-view neighbors to form geometrically grounded representations. A Perceiver-style text-aware readout over frozen CLIP visual features replaces global language conditioning with instruction-relevant evidence retrieval. During training only, a pretrained depth teacher provides supervision for the depth-distribution head; raw depth remains the policy input at training and inference, so there is no test-time teacher overhead.

On RoboTwin 2.0, PEAfowl improves the strongest baseline by 23.0 percentage points in the domain-randomized setting. Physical experiments further show improved performance on the evaluated real-robot tasks.

Architecture

PEAfowl architecture

Geometry-guided multi-view fusion, language-guided perception, and diffusion action prediction.

Evaluation

Simulation

RoboTwin 2.0, 9 tasks, 50 demonstrations per task, and 100 rollouts per task and setting.

  • Clean: PEAfowl 69.6% vs. SEM 51.0%.
  • Domain-randomized: PEAfowl 47.1% vs. SEM 24.1%.

Real robot

Dual-arm AgileX Piper with four Intel RealSense D435 RGB-D cameras; 6 evaluated tasks and 10 trials per task.

  • PEAfowl: 41/60 (68.3%).
  • Without depth distillation: 22/60 (36.7%).
  • SEM: 7/60 (11.7%).
Simulation and real-world setups

Simulation and real-world experimental setups.

Qualitative analysis

Cross-view token consistency

Compared with SEM and pre-aggregation features, PEAfowl after cross-view aggregation produces more coherent representations for corresponding regions across cameras.

Cross-view token consistency

Depth-distribution predictions

The depth-distribution head yields sharper and more complete geometric cues under noisy and incomplete commodity depth, with distillation used only during training.

Depth-distribution predictions

Citation

Please cite the IEEE RA-L paper using its DOI.

@article{fan2026peafowl,
  title   = {PEAfowl: Perception-Enhanced Multi-View Vision-Language-Action for Bimanual Manipulation},
  author  = {Fan, Qingyu and Li, Zhaoxiang and Hu, Jinrui and Lu, Yi and Chen, Wang and Shen, Qiu and Long, Xiao-xiao and Cai, Yinghao and Lu, Tao and Wang, Shuo and Cao, Xun},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2026},
  doi     = {10.1109/LRA.2026.3726379}
}