Selected work

Representative Papers

Swim2Real: VLM-Guided System Identification for Sim-to-Real Transfer

Mar 21, 2026

This work addresses the challenge of system identification and simulation calibration for soft underwater robots, which is hindered by strong nonlinear fluid–structure interactions and the sim-to-real gap. The authors propose the first use of a vision-language model (VLM) for underwater robot system identification, enabling end-to-end calibration of a 16-parameter fish-like robot simulator by directly comparing real swimming videos with simulation outputs—without requiring hand-designed search strategies. By integrating backtracking line search to improve parameter update acceptance rates, the method facilitates zero-shot transfer of reinforcement learning policies from simulation to the physical robot. After calibration, the mean absolute error (MAE) in swimming speed drops to 7.4 mm/s, a 43% improvement over the next-best method, with consistent convergence across five trials. Downstream RL policies achieve 12% and 90% greater swimming distances on the real robot compared to BayesOpt and CMA-ES baselines, respectively.

0 citationsRead paper

Reward-Conditioned Reinforcement Learning

Mar 05, 2026

Traditional reinforcement learning relies on fixed reward functions, limiting its adaptability to changing preferences and generalization capabilities. This work proposes Reward-Conditioned Reinforcement Learning (RCRL), a framework that conditions the policy on reward parameters, enabling off-policy learning of optimal policies for an entire family of reward objectives from experience collected under a single nominal target. By sharing a unified replay buffer, RCRL achieves, for the first time, efficient support for multiple reward goals within a single policy, combining the simplicity of single-task training with the flexibility of multi-task adaptation. Experimental results demonstrate that RCRL consistently outperforms existing baselines across single-task, multi-task, and visual benchmark settings, not only improving performance under the nominal reward but also enabling rapid generalization to new reward parameters.

0 citationsRead paper

Vid2Sid: Videos Can Help Close the Sim2Real Gap

Feb 22, 2026

This work addresses the challenge of accurately calibrating physics parameters in robotic simulators to real hardware when only external camera videos are available and direct force or state measurements are absent. To this end, the authors propose Vid2Sid, the first closed-loop optimization framework that integrates vision-language models (VLMs) into system identification. By analyzing paired simulation-to-reality videos, Vid2Sid leverages foundation vision models and VLMs to perform interpretable semantic reasoning, diagnosing physical mismatches and iteratively refining parameters such as friction, damping, and stiffness. Evaluated on both rigid-body (MuJoCo) and soft-body (PyElastica) systems, Vid2Sid achieves simulation-to-simulation parameter recovery errors below 13%—substantially outperforming baselines (28–98%)—and demonstrates superior average performance on unseen sim-to-real control tasks compared to conventional black-box optimization approaches.

0 citationsRead paper

Debate2Create: Robot Co-design via Large Language Model Debates

Oct 29, 2025

Robot morphology and control co-design faces challenges of an exponentially large search space and strong morphology–behavior coupling. Method: This paper proposes a structured dialectical debate framework powered by large language model (LLM)-based agents: a morphology agent generates structural improvements; a control agent concurrently customizes reward functions; and a diverse panel of evaluators assesses designs via physics simulation, providing feedback to drive closed-loop iterative optimization. Contribution/Results: To our knowledge, this is the first work to introduce multi-agent dialectical reasoning into robotic co-evolution—enabling spontaneous emergence of diverse, specialized morphologies without explicit diversity constraints. Evaluated on quadrupedal locomotion, the resulting robots achieve a 73% average improvement in forward travel distance over baseline methods, demonstrating both efficacy and novelty in automated robot design.

0 citationsRead paper

Decoupled Relative Learning Rate Schedules

Jul 04, 2025

Conventional Transformer training employs a uniform learning rate across all model components, ignoring their dynamic heterogeneity in parameter sensitivity and update magnitude—leading to suboptimal optimization efficiency. Method: We propose a dynamic decoupled learning rate scheduling framework, introducing— for the first time—the concept of *relative learning rates*, which adaptively scale per-component learning rates based on layer-specific gradient statistics and architectural roles. Our approach is architecture-agnostic within the Transformer family and integrates seamlessly with Mixture of Experts (MoE) configurations. Contribution/Results: The method enables direct hyperparameter transfer across model scales—from small baselines to models 27× larger—without manual retuning. Empirical evaluation demonstrates up to 23% faster convergence for complex models and substantial reductions in computational resource consumption. This work establishes a scalable, efficient, and broadly generalizable optimization paradigm for large-scale neural networks.

0 citationsRead paper

Recent publications

Latest Papers

Swim2Real: VLM-Guided System Identification for Sim-to-Real Transfer

Mar 21, 2026

This work addresses the challenge of system identification and simulation calibration for soft underwater robots, which is hindered by strong nonlinear fluid–structure interactions and the sim-to-real gap. The authors propose the first use of a vision-language model (VLM) for underwater robot system identification, enabling end-to-end calibration of a 16-parameter fish-like robot simulator by directly comparing real swimming videos with simulation outputs—without requiring hand-designed search strategies. By integrating backtracking line search to improve parameter update acceptance rates, the method facilitates zero-shot transfer of reinforcement learning policies from simulation to the physical robot. After calibration, the mean absolute error (MAE) in swimming speed drops to 7.4 mm/s, a 43% improvement over the next-best method, with consistent convergence across five trials. Downstream RL policies achieve 12% and 90% greater swimming distances on the real robot compared to BayesOpt and CMA-ES baselines, respectively.

0 citationsRead paper

Reward-Conditioned Reinforcement Learning

Mar 05, 2026

Traditional reinforcement learning relies on fixed reward functions, limiting its adaptability to changing preferences and generalization capabilities. This work proposes Reward-Conditioned Reinforcement Learning (RCRL), a framework that conditions the policy on reward parameters, enabling off-policy learning of optimal policies for an entire family of reward objectives from experience collected under a single nominal target. By sharing a unified replay buffer, RCRL achieves, for the first time, efficient support for multiple reward goals within a single policy, combining the simplicity of single-task training with the flexibility of multi-task adaptation. Experimental results demonstrate that RCRL consistently outperforms existing baselines across single-task, multi-task, and visual benchmark settings, not only improving performance under the nominal reward but also enabling rapid generalization to new reward parameters.

0 citationsRead paper

Vid2Sid: Videos Can Help Close the Sim2Real Gap

Feb 22, 2026

This work addresses the challenge of accurately calibrating physics parameters in robotic simulators to real hardware when only external camera videos are available and direct force or state measurements are absent. To this end, the authors propose Vid2Sid, the first closed-loop optimization framework that integrates vision-language models (VLMs) into system identification. By analyzing paired simulation-to-reality videos, Vid2Sid leverages foundation vision models and VLMs to perform interpretable semantic reasoning, diagnosing physical mismatches and iteratively refining parameters such as friction, damping, and stiffness. Evaluated on both rigid-body (MuJoCo) and soft-body (PyElastica) systems, Vid2Sid achieves simulation-to-simulation parameter recovery errors below 13%—substantially outperforming baselines (28–98%)—and demonstrates superior average performance on unseen sim-to-real control tasks compared to conventional black-box optimization approaches.

0 citationsRead paper

Debate2Create: Robot Co-design via Large Language Model Debates

Oct 29, 2025

Robot morphology and control co-design faces challenges of an exponentially large search space and strong morphology–behavior coupling. Method: This paper proposes a structured dialectical debate framework powered by large language model (LLM)-based agents: a morphology agent generates structural improvements; a control agent concurrently customizes reward functions; and a diverse panel of evaluators assesses designs via physics simulation, providing feedback to drive closed-loop iterative optimization. Contribution/Results: To our knowledge, this is the first work to introduce multi-agent dialectical reasoning into robotic co-evolution—enabling spontaneous emergence of diverse, specialized morphologies without explicit diversity constraints. Evaluated on quadrupedal locomotion, the resulting robots achieve a 73% average improvement in forward travel distance over baseline methods, demonstrating both efficacy and novelty in automated robot design.

0 citationsRead paper

Decoupled Relative Learning Rate Schedules

Jul 04, 2025

Conventional Transformer training employs a uniform learning rate across all model components, ignoring their dynamic heterogeneity in parameter sensitivity and update magnitude—leading to suboptimal optimization efficiency. Method: We propose a dynamic decoupled learning rate scheduling framework, introducing— for the first time—the concept of *relative learning rates*, which adaptively scale per-component learning rates based on layer-specific gradient statistics and architectural roles. Our approach is architecture-agnostic within the Transformer family and integrates seamlessly with Mixture of Experts (MoE) configurations. Contribution/Results: The method enables direct hyperparameter transfer across model scales—from small baselines to models 27× larger—without manual retuning. Empirical evaluation demonstrates up to 23% faster convergence for complex models and substantial reductions in computational resource consumption. This work establishes a scalable, efficient, and broadly generalizable optimization paradigm for large-scale neural networks.

0 citationsRead paper