Selected work

Representative Papers

TrustRoboReward: Preference-Ordered Isotonic Score Editing for Multi-Paradigm Robot Reward Models

Aug 09, 2026

Existing robotic reward models struggle to simultaneously maintain pointwise scoring accuracy and pairwise preference consistency in long-horizon tasks, leading to training noise and performance degradation. This work proposes Preference-Ordered Isotonic Score Editing (POISE), a novel method that achieves conflict-free alignment between these two signal types for the first time, effectively resolving the score-preference reversal problem. Evaluated on a unified four-paradigm dataset and leveraging a vision-language model with video question-answering supervision and the TrustJudge reasoning aggregation strategy, Qwen3-VL-4B calibrated by POISE attains a reward accuracy of 77.96% and improves score-preference consistency to 71.90%. Further integration of TrustJudge elevates the overall performance to 78.57%, surpassing the teacher model.

0 citationsRead paper

Pelican-VLA 0.5: Attending Before Acting Benefits Generalization

Jul 07, 2026

This work addresses the challenge of enabling vision-language-action (VLA) models to attend to instruction-relevant objects and contact regions in unseen environments and robot morphologies, without object annotations, segmentation masks, or task-specific fine-tuning. The authors propose a unified VLA architecture that integrates visual-language understanding, future frame prediction, and action generation, featuring learnable reasoning slots that serve as a compact information bottleneck between perception and action to efficiently route task-relevant signals. During pretraining, the model develops an operation-centric attention hierarchy that generalizes across tasks, allowing it to adapt to diverse policy structures without additional supervision. Experiments demonstrate that the method significantly outperforms existing open-source VLA baselines in zero-shot settings, exhibiting strong cross-scenario and cross-morphology generalization as well as precise attention to manipulation-relevant regions.

0 citationsRead paper

Pelican-VL 1.0: A Foundation Brain Model for Embodied Intelligence

Oct 30, 2025

Embodied intelligence faces challenges in environment perception and adaptive decision-making. Method: This work introduces an open-source multimodal “foundation brain model” architecture, scalable from 7B to 72B parameters and deployable across diverse physical embodiments. To address perceptual and behavioral adaptation, we propose DPPO (Deep-Reflection Proficient Policy Optimization), a meta-cyclic training framework integrating Reinforcement Learning, Refinement, Diagnostic Feedback, and Supervised Fine-Tuning—mimicking human metacognition for efficient deliberate practice. Trained on >4B high-quality tokens using an A800 cluster, the method closes the RL loop with diagnostic feedback and SFT. Contribution/Results: Our approach achieves a 20.3% improvement over base models, outperforms open-source models exceeding 100B parameters by 10.6%, and attains state-of-the-art performance on major embodied AI benchmarks—matching or exceeding closed-source SOTA systems.

0 citationsRead paper

OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation

Aug 16, 2025

Vision-based motion policies often suffer from poor generalization due to over-reliance on fixed viewpoints (e.g., camera pose and background), hindering robust 3D representation learning via multi-view fusion. To address this, we propose OmniD—a BEV (bird’s-eye view) representation learning framework grounded in multi-view image fusion. Its core is the Omni-Feature Generator, which employs deformable attention to enable task-driven feature selection and efficient cross-view feature aggregation. Integrated with diffusion-based policy learning and BEV projection, OmniD produces compact, semantically rich spatial representations that effectively suppress viewpoint-specific noise and background interference. Extensive experiments demonstrate OmniD’s strong generalization: it achieves average improvements of 11%, 17%, and 84% over state-of-the-art methods on in-distribution, out-of-distribution, and few-shot transfer benchmarks, respectively.

0 citationsRead paper

FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency

Jun 10, 2025

Existing generative visuomotor policies rely on multi-step sampling, resulting in high inference latency and failing to meet real-time robotic manipulation requirements. The core challenge lies in enforcing strong temporal continuity and structural consistency in action trajectories—properties that image-acceleration techniques cannot directly transfer. This paper proposes the first frequency-consistent modeling paradigm for streaming visuomotor policies, innovatively introducing temporal frequency-domain consistency constraints and an adaptive consistency loss to explicitly model dynamic trajectory continuity. Our method builds upon invertible flow models, integrating frequency-domain feature alignment, adaptive weighted loss, and end-to-end Vision-Language-Action (VLA) integration. Evaluated on 53 simulated tasks, it surpasses state-of-the-art single-step action generators. When integrated into a VLA framework, it achieves inference acceleration on Libero-40 with zero performance degradation. On physical hardware, it operates at 93.5 Hz.

0 citationsRead paper

Recent publications

Latest Papers

TrustRoboReward: Preference-Ordered Isotonic Score Editing for Multi-Paradigm Robot Reward Models

Aug 09, 2026

Existing robotic reward models struggle to simultaneously maintain pointwise scoring accuracy and pairwise preference consistency in long-horizon tasks, leading to training noise and performance degradation. This work proposes Preference-Ordered Isotonic Score Editing (POISE), a novel method that achieves conflict-free alignment between these two signal types for the first time, effectively resolving the score-preference reversal problem. Evaluated on a unified four-paradigm dataset and leveraging a vision-language model with video question-answering supervision and the TrustJudge reasoning aggregation strategy, Qwen3-VL-4B calibrated by POISE attains a reward accuracy of 77.96% and improves score-preference consistency to 71.90%. Further integration of TrustJudge elevates the overall performance to 78.57%, surpassing the teacher model.

0 citationsRead paper

Pelican-VLA 0.5: Attending Before Acting Benefits Generalization

Jul 07, 2026

This work addresses the challenge of enabling vision-language-action (VLA) models to attend to instruction-relevant objects and contact regions in unseen environments and robot morphologies, without object annotations, segmentation masks, or task-specific fine-tuning. The authors propose a unified VLA architecture that integrates visual-language understanding, future frame prediction, and action generation, featuring learnable reasoning slots that serve as a compact information bottleneck between perception and action to efficiently route task-relevant signals. During pretraining, the model develops an operation-centric attention hierarchy that generalizes across tasks, allowing it to adapt to diverse policy structures without additional supervision. Experiments demonstrate that the method significantly outperforms existing open-source VLA baselines in zero-shot settings, exhibiting strong cross-scenario and cross-morphology generalization as well as precise attention to manipulation-relevant regions.

0 citationsRead paper

Pelican-VL 1.0: A Foundation Brain Model for Embodied Intelligence

Oct 30, 2025

Embodied intelligence faces challenges in environment perception and adaptive decision-making. Method: This work introduces an open-source multimodal “foundation brain model” architecture, scalable from 7B to 72B parameters and deployable across diverse physical embodiments. To address perceptual and behavioral adaptation, we propose DPPO (Deep-Reflection Proficient Policy Optimization), a meta-cyclic training framework integrating Reinforcement Learning, Refinement, Diagnostic Feedback, and Supervised Fine-Tuning—mimicking human metacognition for efficient deliberate practice. Trained on >4B high-quality tokens using an A800 cluster, the method closes the RL loop with diagnostic feedback and SFT. Contribution/Results: Our approach achieves a 20.3% improvement over base models, outperforms open-source models exceeding 100B parameters by 10.6%, and attains state-of-the-art performance on major embodied AI benchmarks—matching or exceeding closed-source SOTA systems.

0 citationsRead paper

OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation

Aug 16, 2025

Vision-based motion policies often suffer from poor generalization due to over-reliance on fixed viewpoints (e.g., camera pose and background), hindering robust 3D representation learning via multi-view fusion. To address this, we propose OmniD—a BEV (bird’s-eye view) representation learning framework grounded in multi-view image fusion. Its core is the Omni-Feature Generator, which employs deformable attention to enable task-driven feature selection and efficient cross-view feature aggregation. Integrated with diffusion-based policy learning and BEV projection, OmniD produces compact, semantically rich spatial representations that effectively suppress viewpoint-specific noise and background interference. Extensive experiments demonstrate OmniD’s strong generalization: it achieves average improvements of 11%, 17%, and 84% over state-of-the-art methods on in-distribution, out-of-distribution, and few-shot transfer benchmarks, respectively.

0 citationsRead paper

FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency

Jun 10, 2025

Existing generative visuomotor policies rely on multi-step sampling, resulting in high inference latency and failing to meet real-time robotic manipulation requirements. The core challenge lies in enforcing strong temporal continuity and structural consistency in action trajectories—properties that image-acceleration techniques cannot directly transfer. This paper proposes the first frequency-consistent modeling paradigm for streaming visuomotor policies, innovatively introducing temporal frequency-domain consistency constraints and an adaptive consistency loss to explicitly model dynamic trajectory continuity. Our method builds upon invertible flow models, integrating frequency-domain feature alignment, adaptive weighted loss, and end-to-end Vision-Language-Action (VLA) integration. Evaluated on 53 simulated tasks, it surpasses state-of-the-art single-step action generators. When integrated into a VLA framework, it achieves inference acceleration on Libero-40 with zero performance degradation. On physical hardware, it operates at 93.5 Hz.

0 citationsRead paper