RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments
研究提出RSIAgent框架，通过自主构建记忆和多代理协调，使数字代理无需训练即可适应新环境，解决预训练模型不足的问题。
0 citationsRead paper
Selected work
Representative Papers
Causally Debiased Latent Action Model for Embodied Action Conditioned World Models
Existing latent action models often entangle action-irrelevant factors—such as background—into action representations when trained on unlabeled videos, compromising controllability. This work proposes CD-LAM, a causal debiasing framework that explicitly formulates the action-irrelevant bias problem and introduces three lightweight fine-tuning objectives: embodied reconstruction, action contrastive learning, and latent space calibration, to learn disentangled and controllable action representations. We further introduce quantifiable evaluation metrics and demonstrate on 2B/14B-parameter models that CD-LAM achieves substantial improvements in action following fidelity, visual quality, and downstream task adaptation efficiency with only 6k fine-tuning steps, reducing the number of updates required for robotic action adaptation by over 12×.
0 citationsRead paper
Recent publications
Latest Papers
RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments
研究提出RSIAgent框架，通过自主构建记忆和多代理协调，使数字代理无需训练即可适应新环境，解决预训练模型不足的问题。
0 citationsRead paper
Causally Debiased Latent Action Model for Embodied Action Conditioned World Models
Existing latent action models often entangle action-irrelevant factors—such as background—into action representations when trained on unlabeled videos, compromising controllability. This work proposes CD-LAM, a causal debiasing framework that explicitly formulates the action-irrelevant bias problem and introduces three lightweight fine-tuning objectives: embodied reconstruction, action contrastive learning, and latent space calibration, to learn disentangled and controllable action representations. We further introduce quantifiable evaluation metrics and demonstrate on 2B/14B-parameter models that CD-LAM achieves substantial improvements in action following fidelity, visual quality, and downstream task adaptation efficiency with only 6k fine-tuning steps, reducing the number of updates required for robotic action adaptation by over 12×.
0 citationsRead paper