arXiv subjects

Sicheng He

Publications and source records attributed to Sicheng He.

At least 19 recordsLinked to original sources

arXiv2026

SurGE: Surrogate Gradient-guided Evolution for Co-design of Legged Robots with Parallel Elasticity

Co-design of legged robots with elastic elements is challenging due to the non-differentiability of contact dynamics and mechanism engagement. This paper presents SurGE, a framework that computes surrogate gradients of the design objective through a differentiable pipeline consisting of a kinodynamic single-rigid-body (Kino-SRB) model and a design-aware control policy, and injects them into CMA-ES via mean shift with cosine-annealed step decay. On a 4-DOF design space of a hopping robot with unidirectional parallel spring, SurGE achieves 6 times lower cross-seed standard deviation and 18% tighter population concentration compared to vanilla CMA-ES, while matching or improving the best objective. Hardware experiments on a 2D design subspace show that, starting from a hand-tuned initial design, SurGE reduces the design objective by 37.65% on hardware, with the improvement trend identified in simulation transferring consistently to the physical system. SurGE provides the potential to accelerate non-differentiable co-design problems in legged robots via surrogate model gradients.

cs.RO↗

arXiv2026

Buffet Alleviation via Linear Stability Adjoint

Transonic buffet, self--sustained shock and shear--layer oscillations, imposes hard limits on the cruise envelope of modern transport aircraft, and avoiding it is a primary design driver. State-of-the-art buffet-onset criteria used in design, such as the $\Delta\alpha = 0.1^\circ$ criterion and separation--sensor methods, are empirical surrogates rather than first--principle predictors, and can yield either overly conservative or unsafe designs. Linear stability analysis (LST) predicts buffet onset directly from the spectrum of the linearized operator about the steady base flow, but using it as an aerodynamic shape optimization constraint has been bottlenecked by the cost of differentiating an eigenvalue with respect to many design variables. In this paper, we develop a coupled adjoint method that efficiently computes the sensitivity of the dominant LST eigenvalue with respect to a large number of shape design variables, by reusing the steady CFD adjoint within a top and bottom level decomposition of the eigenproblem. We verify the eigensolver and adjoint against the canonical cylinder vortex--shedding benchmark, then verify the LST predictions on the OAT15A supercritical airfoil at $M=0.73$, $Re=3.2\times 10^{6}$ against published eigenspectra and against the linear growth phase of a URANS run. Using the resulting gradients, a single-point buffet-constrained drag minimization of the OAT15A achieves a $22.4\%$ drag reduction while satisfying the LST-based buffet constraint. Finally, we present preliminary three-dimensional results on the wing only NASA common research model (CRM) at $M=0.85$, $Re=5\times 10^{6}$, recovering buffet onset at $\alpha \approx 4.0^\circ$ from a sweep of warm--started URANS runs and providing a stepping stone toward three-dimensional buffet-constrained wing optimization with the present adjoint.

physics.flu-dyn↗

arXiv2026

Fast SAM 3D Body: Accelerating SAM 3D Body for Real-Time Full-Body Human Mesh Recovery

SAM 3D Body (3DB) achieves state-of-the-art accuracy in monocular 3D human mesh recovery, yet its inference latency of several seconds per image precludes real-time application. We present Fast SAM 3D Body, a training-free acceleration framework that reformulates the 3DB inference pathway to achieve interactive rates. By decoupling serial spatial dependencies and applying architecture-aware pruning, we enable parallelized multi-crop feature extraction and streamlined transformer decoding. Moreover, to extract the joint-level kinematics (SMPL) compatible with existing humanoid control and policy learning frameworks, we replace the iterative mesh fitting with a direct feedforward mapping, accelerating this specific conversion by over 10,000x. Overall, our framework delivers up to a 10.9x end-to-end speedup while maintaining on-par reconstruction fidelity, even surpassing 3DB on benchmarks such as LSPET. We demonstrate its utility by deploying Fast SAM 3D Body in a vision-only teleoperation system that-unlike methods reliant on wearable IMUs-enables real-time humanoid control and the direct collection of manipulation policies from a single RGB stream.

cs.CV↗

arXiv2026

$\Psi_0$: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation

We introduce $\Psi_0$ (Psi-Zero), an open foundation model to address challenging humanoid loco-manipulation tasks. While existing approaches often attempt to address this fundamental problem by co-training on large and diverse human and humanoid data, we argue that this strategy is suboptimal due to the fundamental kinematic and motion disparities between humans and humanoid robots. Therefore, data efficiency and model performance remain unsatisfactory despite the considerable data volume. To address this challenge, \ours\;decouples the learning process to maximize the utility of heterogeneous data sources. Specifically, we propose a staged training paradigm with different learning objectives: First, we autoregressively pre-train a VLM backbone on large-scale egocentric human videos to acquire generalizable visual-action representations. Then, we post-train a flow-based action expert on high-quality humanoid robot data to learn precise robot joint control. Our research further identifies a critical yet often overlooked data recipe: in contrast to approaches that scale with noisy Internet clips or heterogeneous cross-embodiment robot datasets, we demonstrate that pre-training on high-quality egocentric human manipulation data followed by post-training on domain-specific real-world humanoid trajectories yields superior performance. Extensive real-world experiments demonstrate that \ours\ achieves the best performance using only about 800 hours of human video data and 30 hours of real-world robot data, outperforming baselines pre-trained on more than 10$\times$ as much data by over 40\% in overall success rate across multiple tasks. We will open-source the entire ecosystem to the community, including a data processing and training pipeline, a humanoid foundation model, and a real-time action inference engine.

cs.RO↗

arXiv2026

D-REX: Differentiable Real-to-Sim-to-Real Engine for Learning Dexterous Grasping

Simulation provides a cost-effective and flexible platform for data generation and policy learning to develop robotic systems. However, bridging the gap between simulation and real-world dynamics remains a significant challenge, especially in physical parameter identification. In this work, we introduce a real-to-sim-to-real engine that leverages the Gaussian Splat representations to build a differentiable engine, enabling object mass identification from real-world visual observations and robot control signals, while enabling grasping policy learning simultaneously. Through optimizing the mass of the manipulated object, our method automatically builds high-fidelity and physically plausible digital twins. Additionally, we propose a novel approach to train force-aware grasping policies from limited data by transferring feasible human demonstrations into simulated robot demonstrations. Through comprehensive experiments, we demonstrate that our engine achieves accurate and robust performance in mass identification across various object geometries and mass values. Those optimized mass values facilitate force-aware policy learning, achieving superior and high performance in object grasping, effectively reducing the sim-to-real gap.

cs.RO↗

arXiv2026

Time-Spectral Resolvent Analysis For Periodic Dynamical Systems

Traditional resolvent analysis is a powerful framework for identifying the most amplified input-output structures in fluid flows from a stationary base state. Extending this resolvent analysis to periodic base flows poses computational challenges due to quasi-periodic responses and expensive linearization around a time-varying base flow. This work proposes a time-spectral resolvent operator formulated using the time-spectral method and Fourier collocation that operates directly in the time domain. Rather than mapping between truncated Fourier coefficients as in frequency-domain approaches, the proposed operator maps forcing and response envelopes defined on a discrete temporal grid, enabling direct Jacobian evaluation at collocation points without computing Fourier coefficients of the base flow. The time-spectral resolvent achieves spectral convergence and offers simplified implementation that integrates easily with existing scientific computing tools. The time-spectral resolvent method is validated numerically in three examples including the parametrically forced Mathieu oscillator, the autonomous van der Pol oscillator and the complex Ginzburg-Landau partial differential equation to show that the proposed method accurately predicts the maximum energy amplification and optimal response mode when the system is subject to optimal quasi-periodic forcing. The proposed framework provides a foundation for extending resolvent-based analysis and control to high-dimensional periodic dynamical systems.

math.DS↗

arXiv2026

Efficient Adjoint-based Design Optimization with Optimal Control

Multidisciplinary engineering system design typically employs a sequential process, progressing from system dynamics to design variables and control. However, this process is inefficient and may lead to a suboptimal design. We propose formulating the optimal control and multidisciplinary design optimization (MDO) problems as a single problem with linear quadratic regulator (LQR) control. We use the coupled adjoint method to compute the design variable derivatives, which are critical for gradient-based design optimization. The computational cost of the derivative computation using the adjoint method is independent of the number of design variables, making it suitable for large-scale problems. We show that the coupled adjoint can be solved indirectly and more efficiently by solving three smaller adjoint equations that leverage the feedforward structure of the problem. We demonstrate this new approach on two test problems: design optimization of a classic cart-pole problem and the aerodynamic shape of a quadrotor blade. For the quadrotor blade design problem, we reduce the control cost by 10% by optimizing the blade for a specific control task with a slight penalty in steady hovering power consumption.

math.OC↗

arXiv2026

Training-free score-based diffusion for parameter-dependent stochastic dynamical systems

Simulating parameter-dependent stochastic differential equations (SDEs) presents significant computational challenges, as separate high-fidelity simulations are typically required for each parameter value of interest. Despite the success of machine learning methods in learning SDE dynamics, existing approaches either require expensive neural network training for score function estimation or lack the ability to handle continuous parameter dependence. We present a training-free conditional diffusion model framework for learning stochastic flow maps of parameter-dependent SDEs, where both drift and diffusion coefficients depend on physical parameters. The key technical innovation is a joint kernel-weighted Monte Carlo estimator that approximates the conditional score function using trajectory data sampled at discrete parameter values, enabling interpolation across both state space and the continuous parameter domain. Once trained, the resulting generative model produces sample trajectories for any parameter value within the training range without retraining, significantly accelerating parameter studies, uncertainty quantification, and real-time filtering applications. The performance of the proposed approach is demonstrated via three numerical examples of increasing complexity, showing accurate approximation of conditional distributions across varying parameter values.

stat.ML↗

arXiv2026

Modal-Centric Field Inversion via Differentiable Proper Orthogonal Decomposition

Inverse problems in computational physics often require matching high-dimensional spatio-temporal fields, leading to prohibitive computational costs and ill-conditioned optimizations. We introduce modal-centric field inversion (MCFI), a paradigm that reformulates inverse problems in the reduced space of proper orthogonal decomposition (POD) modes rather than the full physical state space. By targeting dominant flow structures instead of point-wise field values, MCFI provides a compact, physically meaningful objective that naturally regularizes the inversion and dramatically reduces computational burden. Central to this framework is the differentiable POD: an adjoint-based method that efficiently computes sensitivities of POD modes with respect to model parameters, enabling gradient-based optimization in the modal space. We demonstrate MCFI on a one and two-dimensional modified viscous Burger's equation, optimizing spatially varying coefficients to match target dynamics through mode-matching. The adjoint formulation achieves computational cost independent of parameter dimension, in contrast to finite-difference approaches that scale linearly. MCFI establishes a foundation for scalable inverse design and model calibration in unsteady, high-dimensional systems.

math.NA↗

arXiv2025

Torus Time-Spectral Method for Quasi-Periodic Problems

Quasi-periodic trajectories with two or more incommensurate frequencies are ubiquitous in nonlinear dynamics, yet the classical Fourier-based time-spectral method is tied to strictly periodic responses. We introduce a torus time-spectral method that lifts the governing equations to an extended angular phase space, applies double-Fourier collocation on the invariant torus, and solves for the state. The formulation exhibits spectral convergence for quasi-periodic problem which we give a rigorous mathematical proof and also verify numerically. We demonstrate the approach on Duffing oscillators and a nonlinear Klein-Gordon system, documenting spectral error decay on the torus and tight agreement with time-accurate integrations while using modest frequency grids. The method extends naturally to higher-dimensional tori and offers a computationally efficient framework for analyzing quasi-periodic phenomena in fluid mechanics, plasma physics, celestial mechanics, and other domains where multi-frequency dynamics arise.

math.NA↗

arXiv2025

Robot Learning from a Physical World Model

We introduce PhysWorld, a framework that enables robot learning from video generation through physical world modeling. Recent video generation models can synthesize photorealistic visual demonstrations from language commands and images, offering a powerful yet underexplored source of training signals for robotics. However, directly retargeting pixel motions from generated videos to robots neglects physics, often resulting in inaccurate manipulations. PhysWorld addresses this limitation by coupling video generation with physical world reconstruction. Given a single image and a task command, our method generates task-conditioned videos and reconstructs the underlying physical world from the videos, and the generated video motions are grounded into physically accurate actions through object-centric residual reinforcement learning with the physical world model. This synergy transforms implicit visual guidance into physically executable robotic trajectories, eliminating the need for real robot data collection and enabling zero-shot generalizable robotic manipulation. Experiments on diverse real-world tasks demonstrate that PhysWorld substantially improves manipulation accuracy compared to previous approaches. Visit \href{https://pointscoder.github.io/PhysWorld_Web/}{the project webpage} for details.

cs.RO↗

arXiv2025

Adjoint-based Hopf-bifurcation Instability Suppression via First Lyapunov Coefficient

Many physical systems exhibit limit cycle oscillations induced by Hopf bifurcations. In aerospace engineering, limit cycle oscillations arise from undesirable Hopf bifurcation phenomena such as aeroelastic flutter and transonic buffet. In some cases, the resulting limit cycle oscillations can themselves be unstable, leading to amplitude divergence or hysteretic transitions that threaten structural integrity and performance. Avoiding such phenomena when performing gradient based design optimization requires a constraint that quantifies the stability of the bifurcations and the derivative of that constraint with respect to the design variables. To capture the local stability of bifurcations, we leverage the first Lyapunov coefficient, which predicts whether the resulting limit cycle oscillation is stable or unstable. We develop an accurate and efficient method for computing derivatives of the first Lyapunov coefficient. We leverage the adjoint method and reverse algorithmic differentiation to efficiently compute the derivative of the first Lyapunov coefficient. We demonstrate the efficacy of the proposed adjoint method in three design optimization problems that suppress unstable bifurcation: an algebraic Hopf bifurcation model, an aeroelastic model of a typical section, and a nonlinear problem based on the complex Ginzburg-Landau partial differential equation. While the current formulation addresses only a single bifurcation mode, the proposed adjoint shows great potential for efficiently handling Hopf bifurcation constraints in large scale nonlinear problems governed by partial differential equations. Its accuracy, versatility and scalability make it a promising tool for aeroelastic and aerodynamic design optimization as well as other engineering problems involving Hopf bifurcation instabilities.

math.DS↗

arXiv2025

SCOOP'D: Learning Mixed-Liquid-Solid Scooping via Sim2Real Generative Policy

Scooping items with tools such as spoons and ladles is common in daily life, ranging from assistive feeding to retrieving items from environmental disaster sites. However, developing a general and autonomous robotic scooping policy is challenging since it requires reasoning about complex tool-object interactions. Furthermore, scooping often involves manipulating deformable objects, such as granular media or liquids, which is challenging due to their infinite-dimensional configuration spaces and complex dynamics. We propose a method, SCOOP'D, which uses simulation from OmniGibson (built on NVIDIA Omniverse) to collect scooping demonstrations using algorithmic procedures that rely on privileged state information. Then, we use generative policies via diffusion to imitate demonstrations from observational input. We directly apply the learned policy in diverse real-world scenarios, testing its performance on various item quantities, item characteristics, and container types. In zero-shot deployment, our method demonstrates promising results across 465 trials in diverse scenarios, including objects of different difficulty levels that we categorize as "Level 1" and "Level 2." SCOOP'D outperforms all baselines and ablations, suggesting that this is a promising approach to acquiring robotic scooping skills. Project page is at https://scoopdiff.github.io/.

cs.RO↗

arXiv2025

GundamQ: Multi-Scale Spatio-Temporal Representation Learning for Robust Robot Path Planning

In dynamic and uncertain environments, robotic path planning demands accurate spatiotemporal environment understanding combined with robust decision-making under partial observability. However, current deep reinforcement learning-based path planning methods face two fundamental limitations: (1) insufficient modeling of multi-scale temporal dependencies, resulting in suboptimal adaptability in dynamic scenarios, and (2) inefficient exploration-exploitation balance, leading to degraded path quality. To address these challenges, we propose GundamQ: A Multi-Scale Spatiotemporal Q-Network for Robotic Path Planning. The framework comprises two key modules: (i) the Spatiotemporal Perception module, which hierarchically extracts multi-granularity spatial features and multi-scale temporal dependencies ranging from instantaneous to extended time horizons, thereby improving perception accuracy in dynamic environments; and (ii) the Adaptive Policy Optimization module, which balances exploration and exploitation during training while optimizing for smoothness and collision probability through constrained policy updates. Experiments in dynamic environments demonstrate that GundamQ achieves a 15.3\% improvement in success rate and a 21.7\% increase in overall path quality, significantly outperforming existing state-of-the-art methods.

cs.RO↗

arXiv2025

DKFNet: Differentiable Kalman Filter for Field Inversion and Machine Learning

The Kalman filter is a fundamental tool for state estimation in dynamical systems. While originally developed for linear Gaussian settings, it has been extended to nonlinear problems through approaches such as the extended and unscented Kalman filters. Despite its broad use, a persistent limitation is that the underlying approximate model is fixed, which can lead to significant deviations from the true system dynamics. To address this limitation, we introduce the differentiable Kalman filter (DKF), an adjoint-based two-level optimization framework designed to reduce the mismatch between approximate and true dynamics. Within this framework, a field inversion step first uncovers the discrepancy, after which a closure model is trained to capture the discovered dynamics, allowing the filter to adapt flexibly and scale efficiently. We illustrate the capabilities of the DKF using two representative examples: a rocket dynamics model and the Allen-Cahn boundary value problem. In both cases, and across a range of noise levels, the DKF consistently reduces state reconstruction error by at least 90% compared to the classical Kalman filter, while also maintaining robust uncertainty quantification. These results demonstrate that the DKF not only improves estimation accuracy by large margins but also enhances interpretability and scalability, offering a principled pathway for combining data assimilation with modern machine learning.

math.OC↗

arXiv2025

UniFoil: A Universal Dataset of Airfoils in Transitional and Turbulent Regimes for Subsonic and Transonic Flows

We present UniFoil, a large publicly available universal airfoil dataset based on Reynolds-averaged Navier-Stokes (RANS) simulations. It contains over 500,000 samples spanning a wide range of Reynolds and Mach numbers, capturing both transitional and fully turbulent flows across incompressible to compressible regimes. UniFoil is designed to support machine learning research in fluid dynamics, particularly for modeling complex aerodynamic phenomena. Most existing datasets are limited to incompressible, fully turbulent flows with smooth field characteristics, overlooking the critical physics of laminar\-turbulent transition and shock\-wave interactions\-features that exhibit strong nonlinearity and sharp gradients. UniFoil addresses this limitation by offering a broad spectrum of realistic flow conditions. Turbulent simulations utilize the Spalart\-Allmaras (SA) model, while transitional flows are modeled using an e^N\-based transition prediction method coupled with the SA model. The dataset includes a comprehensive geometry set comprising over 4,800 natural laminar flow (NLF) airfoils and 30,000 fully turbulent (FT) airfoils, covering a diverse range of airfoil designs relevant to aerospace, wind energy, and marine applications. This dataset is also valuable for scientific machine learning, enabling the development of data-driven models that more accurately capture the transport processes associated with laminar-turbulent transition. UniFoil is freely available under a permissive CC\-BY\-SA license.

physics.flu-dyn↗

arXiv2025

Automated CAD Modeling Sequence Generation from Text Descriptions via Transformer-Based Large Language Models

Designing complex computer-aided design (CAD) models is often time-consuming due to challenges such as computational inefficiency and the difficulty of generating precise models. We propose a novel language-guided framework for industrial design automation to address these issues, integrating large language models (LLMs) with computer-automated design (CAutoD).Through this framework, CAD models are automatically generated from parameters and appearance descriptions, supporting the automation of design tasks during the detailed CAD design phase. Our approach introduces three key innovations: (1) a semi-automated data annotation pipeline that leverages LLMs and vision-language large models (VLLMs) to generate high-quality parameters and appearance descriptions; (2) a Transformer-based CAD generator (TCADGen) that predicts modeling sequences via dual-channel feature aggregation; (3) an enhanced CAD modeling generation model, called CADLLM, that is designed to refine the generated sequences by incorporating the confidence scores from TCADGen. Experimental results demonstrate that the proposed approach outperforms traditional methods in both accuracy and efficiency, providing a powerful tool for automating industrial workflows and generating complex CAD models from textual prompts. The code is available at https://jianxliao.github.io/cadllm-page/

cs.AI↗

arXiv2025

Sequential Multi-Object Grasping with One Dexterous Hand

Sequentially grasping multiple objects with multi-fingered hands is common in daily life, where humans can fully leverage the dexterity of their hands to enclose multiple objects. However, the diversity of object geometries and the complex contact interactions required for high-DOF hands to grasp one object while enclosing another make sequential multi-object grasping challenging for robots. In this paper, we propose SeqMultiGrasp, a system for sequentially grasping objects with a four-fingered Allegro Hand. We focus on sequentially grasping two objects, ensuring that the hand fully encloses one object before lifting it and then grasps the second object without dropping the first. Our system first synthesizes single-object grasp candidates, where each grasp is constrained to use only a subset of the hand's links. These grasps are then validated in a physics simulator to ensure stability and feasibility. Next, we merge the validated single-object grasp poses to construct multi-object grasp configurations. For real-world deployment, we train a diffusion model conditioned on point clouds to propose grasp poses, followed by a heuristic-based execution strategy. We test our system using $8 \times 8$ object combinations in simulation and $6 \times 3$ object combinations in real. Our diffusion-based grasp model obtains an average success rate of 65.8% over 1,600 simulation trials and 56.7% over 90 real-world trials, suggesting that it is a promising approach for sequential multi-object grasping with multi-fingered hands. Supplementary material is available on our project website: https://hesic73.github.io/SeqMultiGrasp.

cs.RO↗