arXiv subjects

James Zhu

Publications and source records attributed to James Zhu.

11 recordsLinked to original sources

arXiv2026

Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

Agent harnesses (the system prompt, tool set, execution hooks, and context-management scaffolding around a model) are a critical determinant of agentic task success. Automated harness evolution can enable smaller models to perform well on domain-specific tasks at a fraction of frontier-model cost. Since both the harness and model weights shape behavior, we ask how harness evolution and lightweight fine-tuning should be combined. Across seven enterprise agent tasks, we first evolve a harness with the weaker model, then find that a stronger expert often uses it more effectively, suggesting expert supervision could close the remaining gap. However, training the weaker model on the expert's complete trajectories under the evolved harness backfires: performance regresses on all seven tasks by 4 to 30 points across Qwen3-Coder and Gemma 4, even though the same procedure helps under the unevolved harness. Our analysis shows that imitation transfers knowledge and increases scaffold usage, but disrupts model-harness fit: the weaker model adopts the expert's planning strategy without the competence to execute it and no longer matches the harness evolved around its native planning style. We therefore develop an on-policy expert-correction pipeline, automated by a meta-level MLE agent, that localizes the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn. This preserves the model's planning style and combines the gains of harness evolution and model adaptation. Our results identify and resolve a source of contention between harness and weight updates, yielding a compatibility-preserving recipe for economical co-evolution on domain-specific enterprise tasks.

cs.AI↗

arXiv2026

DarwinX: Evolving Agent Harnesses Through Natural Selection

An LLM agent's capability depends not only on model weights but on its harness: prompts, tools, skills, and control flow. Self-improvement loops already edit harnesses, yet single-lineage search is path-dependent and local wins often regress other tasks. We introduce DarwinX, which treats self-evolution as selection over a population of harnesses with the model frozen: a preserve-and-extend contract admits only variants that extend coverage without regressing, an archive keeps alternative lineages for recombination, and failure-, teacher-, and self-derived evidence share one edit interface. Fitness comes from each benchmark's own verifier: no gold solutions, no hand-picked winners. Across four benchmarks that progressively separate the evolution signal from the test, one loop adds about 17 points on average: Terminal-Bench 2.1 rises +7.7 to 83.2% on a matched base and to the verified frontier at 84.7% on a stronger one; TerminalWorld's held-out split reaches 68.3%, ahead of every off-the-shelf agent; WebArena-Infinity real-task pass@1 rises from 43.5% to 93.0% audit-clean; and a Terminal-Bench 2.1 harness transfers unchanged to SWE-bench Verified. What evolves is general agent competence, not benchmark-specific patches, so it survives changes of task, verifier, and base model. A frozen model need not be a fixed agent: harness selection turns evaluation compute into durable capability.

cs.NE↗

arXiv2025

Asynchronous Distributed Multi-Robot Motion Planning Under Imperfect Communication

This paper addresses the challenge of coordinating multi-robot systems under realistic communication delays using distributed optimization. We focus on consensus ADMM as a scalable framework for generating collision-free, dynamically feasible motion plans in both trajectory optimization and receding-horizon control settings. In practice, however, these algorithms are sensitive to penalty tuning or adaptation schemes (e.g. residual balancing and adaptive parameter heuristics) that do not explicitly consider delays. To address this, we introduce a Delay-Aware ADMM (DA-ADMM) variant that adapts penalty parameters based on real-time delay statistics, allowing agents to down-weight stale information and prioritize recent updates during consensus and dual updates. Through extensive simulations in 2D and 3D environments with double-integrator, Dubins-car, and drone dynamics, we show that DA-ADMM significantly improves robustness, success rate, and solution quality compared to fixed-parameter, residual-balancing, and fixed-constraint baselines. Our results highlight that performance degradation is not solely determined by delay length or frequency, but by the optimizer's ability to contextually reason over delayed information. The proposed DA-ADMM achieves consistently better coordination performance across a wide range of delay conditions, offering a principled and efficient mechanism for resilient multi-robot motion planning under imperfect communication.

cs.RO↗

arXiv2024

Rate, Explain and Cite (REC): Enhanced Explanation and Attribution in Automatic Evaluation by Large Language Models

LLMs have demonstrated impressive proficiency in generating coherent and high-quality text, making them valuable across a range of text-generation tasks. However, rigorous evaluation of this generated content is crucial, as ensuring its quality remains a significant challenge due to persistent issues such as factual inaccuracies and hallucination. This paper introduces three fine-tuned general-purpose LLM autoevaluators, REC-8B, REC-12B and REC-70B, specifically designed to evaluate generated text across several dimensions: faithfulness, instruction following, coherence, and completeness. These models not only provide ratings for these metrics but also offer detailed explanation and verifiable citation, thereby enhancing trust in the content. Moreover, the models support various citation modes, accommodating different requirements for latency and granularity. Extensive evaluations on diverse benchmarks demonstrate that our general-purpose LLM auto-evaluator, REC-70B, outperforms state-of-the-art LLMs, excelling in content evaluation by delivering better quality explanation and citation with minimal bias. Our REC dataset and models are available at https://github.com/adelaidehsu/REC.

cs.CL↗

arXiv2024

Hybrid Iterative Linear Quadratic Estimation: Optimal Estimation for Hybrid Systems

In this paper we present Hybrid iterative Linear Quadratic Estimation (HiLQE), an optimization based offline state estimation algorithm for hybrid dynamical systems. We utilize the saltation matrix, a first order approximation of the variational update through an event driven hybrid transition, to calculate gradient information through hybrid events in the backward pass of an iterative linear quadratic optimization over state estimates. This enables accurate computation of the value function approximation at each timestep. Additionally, the forward pass in the iterative algorithm is augmented with hybrid dynamics in the rollout. A reference extension method is used to account for varying impact times when comparing states for the feedback gain in noise calculation. The proposed method is demonstrated on an ASLIP hopper system with position measurements. In comparison to the Salted Kalman Filter (SKF), the algorithm presented here achieves a maximum of 63.55% reduction in estimation error magnitude over all state dimensions near impact events.

cs.RO↗

arXiv2024

Double-Anonymous Review for Robotics

Prior research has investigated the benefits and costs of double-anonymous review (DAR, also known as double-blind review) in comparison to single-anonymous review (SAR) and open review (OR). Several review papers have attempted to compile experimental results in peer review research both broadly and in engineering and computer science. This document summarizes prior research in peer review that may inform decisions about the format of peer review in the field of robotics and makes some recommendations for potential next steps for robotics publication.

cs.RO↗

arXiv2024

Convergent iLQR for Safe Trajectory Planning and Control of Legged Robots

In order to perform highly dynamic and agile maneuvers, legged robots typically spend time in underactuated domains (e.g. with feet off the ground) where the system has limited command of its acceleration and a constrained amount of time before transitioning to a new domain (e.g. foot touchdown). Meanwhile, these transitions can instantaneously change the system's state, possibly causing perturbations to be mapped arbitrarily far away from the target trajectory. These properties make it difficult for local feedback controllers to effectively recover from disturbances as the system evolves through underactuated domains and hybrid impact events. To address this, we utilize the fundamental solution matrix that characterizes the evolution of perturbations through a hybrid trajectory and its 2-norm, which represents the worst-case growth of perturbations. In this paper, the worst-case perturbation analysis is used to explicitly reason about the tracking performance of a hybrid trajectory and is incorporated in an iLQR framework to optimize a trajectory while taking into account the closed-loop convergence of the trajectory under an LQR tracking controller. The generated convergent trajectories recover more effectively from perturbations, are more robust to large disturbances, and use less feedback control effort than trajectories generated with traditional methods.

cs.RO↗

arXiv2023

Synergistic Perception and Control Simplex for Verifiable Safe Vertical Landing

Perception, Planning, and Control form the essential components of autonomy in advanced air mobility. This work advances the holistic integration of these components to enhance the performance and robustness of the complete cyber-physical system. We adapt Perception Simplex, a system for verifiable collision avoidance amidst obstacle detection faults, to the vertical landing maneuver for autonomous air mobility vehicles. We improve upon this system by replacing static assumptions of control capabilities with dynamic confirmation, i.e., real-time confirmation of control limitations of the system, ensuring reliable fulfillment of safety maneuvers and overrides, without dependence on overly pessimistic assumptions. Parameters defining control system capabilities and limitations, e.g., maximum deceleration, are continuously tracked within the system and used to make safety-critical decisions. We apply these techniques to propose a verifiable collision avoidance solution for autonomous aerial mobility vehicles operating in cluttered and potentially unsafe environments.

cs.RO↗

arXiv2023

Grounding Robot Navigation in Self-Defense Law

Robots operating in close proximity to humans rely heavily on human trust to successfully complete their tasks. But what are the real outcomes when this trust is violated? Self-defense law provides a framework for analyzing tangible failure scenarios that can inform the design of robots and their algorithms. Studying self-defense is particularly important for ground robots since they operate within public environments, where they can pose a legitimate threat to the safety of nearby humans. Moreover, even if ground robots can guarantee human safety, the perception of a physical threat is sufficient to justify human self-defense against robots. In this paper, we synthesize works in law, engineering, and social science to present four actionable recommendations for how the robotics community can craft robots to mitigate the likelihood of self-defense situations arising. We establish how current U.S. self-defense law can justify a human protecting themselves against a robot, discuss the current literature on human attitudes toward robots, and analyze methods that have been produced to allow robots to operate close to humans. Finally, we present hypothetical scenarios that underscore how current robot navigation methods can fail to sufficiently consider self-defense concerns and the need for the recommendations to guide improvements in the field.

cs.RO↗

arXiv2023

Saltation Matrices: The Essential Tool for Linearizing Hybrid Dynamical Systems

Hybrid dynamical systems, i.e. systems that have both continuous and discrete states, are ubiquitous in engineering, but are difficult to work with due to their discontinuous transitions. For example, a robot leg is able to exert very little control effort while it is in the air compared to when it is on the ground. When the leg hits the ground, the penetrating velocity instantaneously collapses to zero. These instantaneous changes in dynamics and discontinuities (or jumps) in state make standard smooth tools for planning, estimation, control, and learning difficult for hybrid systems. One of the key tools for accounting for these jumps is called the saltation matrix. The saltation matrix is the sensitivity update when a hybrid jump occurs and has been used in a variety of fields including robotics, power circuits, and computational neuroscience. This paper presents an intuitive derivation of the saltation matrix and discusses what it captures, where it has been used in the past, how it is used for linear and quadratic forms, how it is computed for rigid body systems with unilateral constraints, and some of the structural properties of the saltation matrix in these cases.

cs.RO↗

arXiv2022

Hybrid Event Shaping to Stabilize Periodic Hybrid Orbits

Many controllers for legged robotic systems leverage open- or closed-loop control at discrete hybrid events to enhance stability. These controllers appear in several well studied phenomena such as the Raibert stepping controller, paddle juggling and swing leg retraction. This work introduces hybrid event shaping (HES): a generalized method for analyzing and producing stable hybrid event controllers. HES utilizes the saltation matrix, which gives a closed-form equation for the effect that hybrid events have on stability. We also introduce shape parameters, which are higher order terms that can be tuned completely independently from the system dynamics to promote stability. Optimization methods are used to produce values of these parameters that optimize a stability measure. Hybrid event shaping captures previously developed control methods while also producing new optimally stable trajectories without the need for continuous-domain feedback.

cs.RO↗