Selected work

Representative Papers

Learning Loco-Manipulation From SMPC Demonstrations With Sparse Offline-to-Online RL

Aug 12, 2026

This work addresses the challenge of scaling reinforcement learning to complex loco-manipulation tasks, where conventional approaches rely heavily on handcrafted dense rewards. The authors propose a novel framework that leverages sample-based model predictive control (SMPC) as an automated expert policy generator to efficiently construct large-scale offline datasets in simulation. This dataset is then used within a hierarchical architecture combining offline-to-online reinforcement learning under sparse rewards with a low-level dynamically stable controller. Notably, the method eliminates the need for manual reward engineering and enables agents trained solely with sparse rewards to surpass the performance of the SMPC teacher policy. The approach demonstrates strong empirical results, successfully deploying on both the Spot quadruped and G1 humanoid robots with high performance, robustness, and effective sim-to-real transfer.

0 citationsRead paper

Prior-SG: Task and Prior Driven Region Segmentation for Scene Graphs in Arbitrarily-Structured Environments

Aug 06, 2026

This work addresses the challenge of accurately segmenting semantic regions in open-ended environments for scene graph construction, a task where existing methods falter due to their reliance on local visual clustering or rigid geometric heuristics. The proposed approach formulates scene graph generation as a probabilistic alignment problem, integrating RGB-D instance maps with dynamic task- and structure-level priors derived from large language models (LLMs). By optimizing a Markov random field that fuses these multimodal sources, the method infers high-level functional semantics. Notably, it introduces task-driven LLM-derived topological priors, enabling zero-shot ontological flexibility and supporting task-adaptive spatial reconceptualization even in the absence of physical partitions. Evaluated in both simulated and real-world open environments, the approach achieves state-of-the-art performance and robustly identifies distant functional boundaries.

0 citationsRead paper

Semantic Homogenization in Italian Popular Music: A Diachronic Analysis

Jul 06, 2026

This study investigates whether semantic diversity in Italian popular music lyrics has declined over time, thereby testing the universality of global semantic convergence beyond English-language contexts. Drawing on 75 years of finalist songs from the Sanremo Music Festival, the authors develop a multi-granular analytical framework that integrates word embeddings and large language models to trace diachronic semantic shifts at the levels of full texts, lyrical segments, thematic content, and individual vocabulary. The research presents the first systematic evidence of progressive semantic convergence in Italian pop lyrics, revealing patterns closely aligned with those observed in English-language songs. These findings not only validate the efficacy of natural language processing techniques for studying long-term cultural expression but also introduce a transferable paradigm for analyzing semantic evolution across linguistic and cultural domains.

0 citationsRead paper

Introducing Background Temperature to Characterise Hidden Randomness in Large Language Models

Apr 24, 2026

This work addresses the often-overlooked output variability in large language models (LLMs) even under nominal temperature settings of \( T = 0 \), which undermines reproducibility and evaluation reliability due to non-determinism inherent in inference implementations. We introduce, for the first time, the concept of “background temperature” (\( T_{\text{bg}} \)) to formally characterize the implicit randomness introduced by the inference environment and establish a mapping between this effect and the equivalent temperature of an idealized reference system. Through stochastic perturbation modeling, an equivalent temperature estimation protocol, and cross-platform experimental validation, we empirically confirm the existence of \( T_{\text{bg}} \) across mainstream LLMs and demonstrate its tangible impact on model deployment and evaluation. This study provides the first quantifiable and measurable theoretical framework for understanding and analyzing inference-induced non-determinism in LLMs.

0 citationsRead paper

ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control

Jan 30, 2026

This work addresses the challenge of deploying agile multi-contact locomotion on humanoid robots, which typically requires extensive skill-specific customization and parameter tuning. The authors propose ZEST, a framework that leverages reinforcement learning to train whole-body control policies end-to-end from diverse motion data—including motion capture, monocular video, and animation—without relying on contact labels, reference windows, state estimators, or intricate reward shaping. By integrating adaptive sampling with a model-based auxiliary torque curriculum, ZEST achieves zero-shot generalization across behaviors and platforms under moderate domain randomization in simulation. Experiments demonstrate successful reproduction of complex multi-contact skills such as crawling and breakdancing on the Atlas robot, direct transfer of dance and box-jumping motions from video to both Atlas and Unitree G1, and even consecutive backflips on the quadrupedal Spot robot, highlighting its strong cross-modal and cross-morphology generalization capabilities.

0 citationsRead paper

Recent publications

Latest Papers

Learning Loco-Manipulation From SMPC Demonstrations With Sparse Offline-to-Online RL

Aug 12, 2026

This work addresses the challenge of scaling reinforcement learning to complex loco-manipulation tasks, where conventional approaches rely heavily on handcrafted dense rewards. The authors propose a novel framework that leverages sample-based model predictive control (SMPC) as an automated expert policy generator to efficiently construct large-scale offline datasets in simulation. This dataset is then used within a hierarchical architecture combining offline-to-online reinforcement learning under sparse rewards with a low-level dynamically stable controller. Notably, the method eliminates the need for manual reward engineering and enables agents trained solely with sparse rewards to surpass the performance of the SMPC teacher policy. The approach demonstrates strong empirical results, successfully deploying on both the Spot quadruped and G1 humanoid robots with high performance, robustness, and effective sim-to-real transfer.

0 citationsRead paper

Prior-SG: Task and Prior Driven Region Segmentation for Scene Graphs in Arbitrarily-Structured Environments

Aug 06, 2026

This work addresses the challenge of accurately segmenting semantic regions in open-ended environments for scene graph construction, a task where existing methods falter due to their reliance on local visual clustering or rigid geometric heuristics. The proposed approach formulates scene graph generation as a probabilistic alignment problem, integrating RGB-D instance maps with dynamic task- and structure-level priors derived from large language models (LLMs). By optimizing a Markov random field that fuses these multimodal sources, the method infers high-level functional semantics. Notably, it introduces task-driven LLM-derived topological priors, enabling zero-shot ontological flexibility and supporting task-adaptive spatial reconceptualization even in the absence of physical partitions. Evaluated in both simulated and real-world open environments, the approach achieves state-of-the-art performance and robustly identifies distant functional boundaries.

0 citationsRead paper

Semantic Homogenization in Italian Popular Music: A Diachronic Analysis

Jul 06, 2026

This study investigates whether semantic diversity in Italian popular music lyrics has declined over time, thereby testing the universality of global semantic convergence beyond English-language contexts. Drawing on 75 years of finalist songs from the Sanremo Music Festival, the authors develop a multi-granular analytical framework that integrates word embeddings and large language models to trace diachronic semantic shifts at the levels of full texts, lyrical segments, thematic content, and individual vocabulary. The research presents the first systematic evidence of progressive semantic convergence in Italian pop lyrics, revealing patterns closely aligned with those observed in English-language songs. These findings not only validate the efficacy of natural language processing techniques for studying long-term cultural expression but also introduce a transferable paradigm for analyzing semantic evolution across linguistic and cultural domains.

0 citationsRead paper

Introducing Background Temperature to Characterise Hidden Randomness in Large Language Models

Apr 24, 2026

This work addresses the often-overlooked output variability in large language models (LLMs) even under nominal temperature settings of \( T = 0 \), which undermines reproducibility and evaluation reliability due to non-determinism inherent in inference implementations. We introduce, for the first time, the concept of “background temperature” (\( T_{\text{bg}} \)) to formally characterize the implicit randomness introduced by the inference environment and establish a mapping between this effect and the equivalent temperature of an idealized reference system. Through stochastic perturbation modeling, an equivalent temperature estimation protocol, and cross-platform experimental validation, we empirically confirm the existence of \( T_{\text{bg}} \) across mainstream LLMs and demonstrate its tangible impact on model deployment and evaluation. This study provides the first quantifiable and measurable theoretical framework for understanding and analyzing inference-induced non-determinism in LLMs.

0 citationsRead paper

ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control

Jan 30, 2026

This work addresses the challenge of deploying agile multi-contact locomotion on humanoid robots, which typically requires extensive skill-specific customization and parameter tuning. The authors propose ZEST, a framework that leverages reinforcement learning to train whole-body control policies end-to-end from diverse motion data—including motion capture, monocular video, and animation—without relying on contact labels, reference windows, state estimators, or intricate reward shaping. By integrating adaptive sampling with a model-based auxiliary torque curriculum, ZEST achieves zero-shot generalization across behaviors and platforms under moderate domain randomization in simulation. Experiments demonstrate successful reproduction of complex multi-contact skills such as crawling and breakdancing on the Atlas robot, direct transfer of dance and box-jumping motions from video to both Atlas and Unitree G1, and even consecutive backflips on the quadrupedal Spot robot, highlighting its strong cross-modal and cross-morphology generalization capabilities.

0 citationsRead paper