XPACE: Joint World and Action Modeling from Heterogeneous Experience
XPACE通过结合异构经验学习和模拟生成，解决了通用机器人选择动作及预测动作后果的问题，提高了行为鲁棒性和任务完成率。
0 citationsRead paper
Selected work
Representative Papers
X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
本文提出X-AuT框架，通过层次选择、表示对齐和跨尺度蒸馏等方法压缩语音大模型的音频编码器，减少推理成本同时保持性能。
0 citationsRead paper
Behavior-Skill: A Fine-Grained Benchmark for Evaluating Vision-Language-Action Policies in Long-Horizon Tasks
为了解决长时任务中多技能执行难题，提出Behavior-Skill基准，通过细粒度技能实例评估视觉-语言-动作策略，提供中间状态恢复和成功条件以独立评测。
0 citationsRead paper
AnyWorld: Factorized Egocentric World Models for Cross-Embodiment Generalization
为解决机器人学习中跨实体泛化问题，提出AnyWorld框架，通过分解动作、视角和实体，将单个人类互动扩展为多样化的机器人原生体验。
0 citationsRead paper
DeepInsight II: One Trace from Benchmark to Robot
This study addresses the fragmentation in embodied intelligence evaluation and the lack of empirical continuity in sim-to-real deployment by proposing a native sim-to-real tracking reduction framework alongside a five-level handover diagnostic system. By integrating MotionBench unified metrics, cross-domain identity binding, and System2-1-0 combinatorial attribution techniques, we establish a shared tracking mechanism between simulation and physical platforms. The research reproduces multi-benchmark checkpoints and achieves trajectory alignment, effectively validating fault attribution and remediation capabilities under hardware constraints. Consequently, this work establishes a complete empirical closed loop from benchmark execution to physical verification, significantly enhancing both the interpretability and deployment reliability of embodied intelligence systems.
0 citationsRead paper
Recent publications
Latest Papers
XPACE: Joint World and Action Modeling from Heterogeneous Experience
XPACE通过结合异构经验学习和模拟生成，解决了通用机器人选择动作及预测动作后果的问题，提高了行为鲁棒性和任务完成率。
0 citationsRead paper
X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
本文提出X-AuT框架，通过层次选择、表示对齐和跨尺度蒸馏等方法压缩语音大模型的音频编码器，减少推理成本同时保持性能。
0 citationsRead paper
Behavior-Skill: A Fine-Grained Benchmark for Evaluating Vision-Language-Action Policies in Long-Horizon Tasks
为了解决长时任务中多技能执行难题，提出Behavior-Skill基准，通过细粒度技能实例评估视觉-语言-动作策略，提供中间状态恢复和成功条件以独立评测。
0 citationsRead paper
AnyWorld: Factorized Egocentric World Models for Cross-Embodiment Generalization
为解决机器人学习中跨实体泛化问题，提出AnyWorld框架，通过分解动作、视角和实体，将单个人类互动扩展为多样化的机器人原生体验。
0 citationsRead paper
DeepInsight II: One Trace from Benchmark to Robot
This study addresses the fragmentation in embodied intelligence evaluation and the lack of empirical continuity in sim-to-real deployment by proposing a native sim-to-real tracking reduction framework alongside a five-level handover diagnostic system. By integrating MotionBench unified metrics, cross-domain identity binding, and System2-1-0 combinatorial attribution techniques, we establish a shared tracking mechanism between simulation and physical platforms. The research reproduces multi-benchmark checkpoints and achieves trajectory alignment, effectively validating fault attribution and remediation capabilities under hardware constraints. Consequently, this work establishes a complete empirical closed loop from benchmark execution to physical verification, significantly enhancing both the interpretability and deployment reliability of embodied intelligence systems.
0 citationsRead paper