Selected work

Representative Papers

Inverse Manipulation through Symbolic Planning and Residual Operator Learning

Jun 03, 2026

This work addresses the challenge of reversing robotic manipulation tasks, which cannot be reliably achieved through symbolic state inversion or trajectory replay alone due to the complexities of continuous dynamics that prevent symbolic inverse planning from accurately reproducing forward execution outcomes. To overcome this limitation, the authors propose a novel paradigm that integrates symbolic inverse planning with residual reinforcement learning. Specifically, soft geometric predicates are automatically extracted from demonstrations to construct STRIPS-style operators and define inverse goals. A task planner then invokes primitive actions to coarsely reverse the task, followed by fine-tuning via a Soft Actor-Critic algorithm that learns a residual policy to precisely satisfy any unmet symbolic predicates. Evaluated on the ManiSkill3 PushCube task, the approach successfully achieves full pose reversal of the cube while maintaining both symbolic logical consistency and physical feasibility.

0 citationsRead paper

Polymander II: an amphibious salamander-inspired robot with contact and flow sensors

May 23, 2026

This work addresses the challenge that existing amphibious robots struggle to simultaneously perceive terrestrial contact forces and underwater hydrodynamic forces, limiting their adaptability and locomotion robustness in complex environments. The authors propose a salamander-inspired amphibious robot featuring an innovative integration of compact Hall-effect sensors, enabling high-frequency synchronous detection of foot-end contact forces and lateral hydrodynamic forces (exteroception >500 Hz, proprioception at 100 Hz). Coupled with a dual-bus communication architecture and joint position–load sensing technology, the system supports multimodal sensor fusion and waterproof embedded operation. Experimental results demonstrate the robot’s efficient traversal across diverse amphibious terrains and highlight its potential to execute complex locomotion tasks through real-time sensory feedback.

0 citationsRead paper

Fast-SmartWay: Panoramic-Free End-to-End Zero-Shot Vision-and-Language Navigation

Nov 02, 2025

Existing zero-shot vision-language navigation (VLN) approaches rely on panoramic observations and two-stage waypoint prediction, resulting in high latency and poor deployability. This paper proposes the first end-to-end framework that eliminates panoramic input, directly predicting actions from only three forward-looking RGB-D frames and natural language instructions. Our method introduces: (1) an uncertainty-aware reasoning mechanism integrating a disambiguation module with future–past bidirectional temporal modeling to enhance decision robustness and long-horizon planning capability; and (2) multimodal large language model (MLLM)-based cross-modal alignment. Evaluated in both simulation and on real robotic platforms, our approach achieves significantly lower per-step latency while matching or surpassing panoramic baseline performance—marking the first demonstration of low-latency, deployable zero-shot VLN.

0 citationsRead paper

Efficient and Encrypted Inference using Binarized Neural Networks within In-Memory Computing Architectures

Oct 27, 2025

Binary neural network (BNN) weights in compute-in-memory (CIM) accelerators are vulnerable to physical extraction, while conventional encryption/decryption undermines CIM’s “compute-as-storage” paradigm. Method: This paper proposes a decryption-free encrypted inference scheme leveraging physically unclonable functions (PUFs) to generate device-unique keys for lightweight encryption of BNN weights; binary inference is performed directly on ciphertexts within the CIM crossbar array—without decryption. Contribution/Results: The approach incurs negligible hardware overhead compared to fully homomorphic encryption, preserving CIM’s energy efficiency and throughput. Experiments show that without the legitimate PUF-derived key, model accuracy collapses to below 15%, effectively thwarting weight theft. Inference latency and energy consumption increase by less than 1%, demonstrating a tight co-optimization of security and efficiency.

0 citationsRead paper

Estimating Peer Effects Using Partial Network Data

Sep 09, 2025

This paper addresses the core challenge in estimating peer effects in social networks—unobserved or mismeasured network structure due to sampling bias, link censoring, or misclassification. Methodologically, it proposes a novel identification strategy that does not require full network data, building instead on a linear mean model framework wherein consistent estimation of the degree distribution suffices for unbiased peer effect identification. This approach mitigates the downward bias inherent in conventional methods that ignore network measurement error. Empirically, using the Add Health dataset, the study demonstrates that correcting for network errors substantially increases estimated peer effects in students’ academic achievement, confirming that neglecting data imperfections systematically underestimates true causal impacts. By relaxing the strong requirement of precise network topology, the paper advances robust and feasible econometric inference for peer effects under partial network observability.

0 citationsRead paper

Recent publications

Latest Papers

Inverse Manipulation through Symbolic Planning and Residual Operator Learning

Jun 03, 2026

This work addresses the challenge of reversing robotic manipulation tasks, which cannot be reliably achieved through symbolic state inversion or trajectory replay alone due to the complexities of continuous dynamics that prevent symbolic inverse planning from accurately reproducing forward execution outcomes. To overcome this limitation, the authors propose a novel paradigm that integrates symbolic inverse planning with residual reinforcement learning. Specifically, soft geometric predicates are automatically extracted from demonstrations to construct STRIPS-style operators and define inverse goals. A task planner then invokes primitive actions to coarsely reverse the task, followed by fine-tuning via a Soft Actor-Critic algorithm that learns a residual policy to precisely satisfy any unmet symbolic predicates. Evaluated on the ManiSkill3 PushCube task, the approach successfully achieves full pose reversal of the cube while maintaining both symbolic logical consistency and physical feasibility.

0 citationsRead paper

Polymander II: an amphibious salamander-inspired robot with contact and flow sensors

May 23, 2026

This work addresses the challenge that existing amphibious robots struggle to simultaneously perceive terrestrial contact forces and underwater hydrodynamic forces, limiting their adaptability and locomotion robustness in complex environments. The authors propose a salamander-inspired amphibious robot featuring an innovative integration of compact Hall-effect sensors, enabling high-frequency synchronous detection of foot-end contact forces and lateral hydrodynamic forces (exteroception >500 Hz, proprioception at 100 Hz). Coupled with a dual-bus communication architecture and joint position–load sensing technology, the system supports multimodal sensor fusion and waterproof embedded operation. Experimental results demonstrate the robot’s efficient traversal across diverse amphibious terrains and highlight its potential to execute complex locomotion tasks through real-time sensory feedback.

0 citationsRead paper

Fast-SmartWay: Panoramic-Free End-to-End Zero-Shot Vision-and-Language Navigation

Nov 02, 2025

Existing zero-shot vision-language navigation (VLN) approaches rely on panoramic observations and two-stage waypoint prediction, resulting in high latency and poor deployability. This paper proposes the first end-to-end framework that eliminates panoramic input, directly predicting actions from only three forward-looking RGB-D frames and natural language instructions. Our method introduces: (1) an uncertainty-aware reasoning mechanism integrating a disambiguation module with future–past bidirectional temporal modeling to enhance decision robustness and long-horizon planning capability; and (2) multimodal large language model (MLLM)-based cross-modal alignment. Evaluated in both simulation and on real robotic platforms, our approach achieves significantly lower per-step latency while matching or surpassing panoramic baseline performance—marking the first demonstration of low-latency, deployable zero-shot VLN.

0 citationsRead paper

Efficient and Encrypted Inference using Binarized Neural Networks within In-Memory Computing Architectures

Oct 27, 2025

Binary neural network (BNN) weights in compute-in-memory (CIM) accelerators are vulnerable to physical extraction, while conventional encryption/decryption undermines CIM’s “compute-as-storage” paradigm. Method: This paper proposes a decryption-free encrypted inference scheme leveraging physically unclonable functions (PUFs) to generate device-unique keys for lightweight encryption of BNN weights; binary inference is performed directly on ciphertexts within the CIM crossbar array—without decryption. Contribution/Results: The approach incurs negligible hardware overhead compared to fully homomorphic encryption, preserving CIM’s energy efficiency and throughput. Experiments show that without the legitimate PUF-derived key, model accuracy collapses to below 15%, effectively thwarting weight theft. Inference latency and energy consumption increase by less than 1%, demonstrating a tight co-optimization of security and efficiency.

0 citationsRead paper

Estimating Peer Effects Using Partial Network Data

Sep 09, 2025

This paper addresses the core challenge in estimating peer effects in social networks—unobserved or mismeasured network structure due to sampling bias, link censoring, or misclassification. Methodologically, it proposes a novel identification strategy that does not require full network data, building instead on a linear mean model framework wherein consistent estimation of the degree distribution suffices for unbiased peer effect identification. This approach mitigates the downward bias inherent in conventional methods that ignore network measurement error. Empirically, using the Add Health dataset, the study demonstrates that correcting for network errors substantially increases estimated peer effects in students’ academic achievement, confirming that neglecting data imperfections systematically underestimates true causal impacts. By relaxing the strong requirement of precise network topology, the paper advances robust and feasible econometric inference for peer effects under partial network observability.

0 citationsRead paper