All papers

Combines real-time digital twins with human-in-the-loop demonstrations to help collaborative robots adapt online using reinforcement learning.

The paper presents instance-anchored interaction evidence to ground robot planning in human pointing and physical object manipulation videos.

The paper proposes RESETTLE, an execution recovery framework triggered by policy proposal disagreements using retrieval and corrective control.

The paper presents MiniWAM, a world action model predicting compact transition targets learned via inverse dynamics and feature reconstruction.

The paper introduces MAP2, combining curvature-based model predictive control with sparse Gaussian process residual learning for autonomous racing.

The paper presents a sim-to-real reinforcement learning pipeline using brief field trajectories and system identification for autonomous surface vehicles.

The paper analyzes action-prediction residuals in robot learning and proposes heteroscedastic Student-t modeling to improve regression policies.

The paper introduces a fixed-reference pose residual model to quantify cross-dataset transfer of human body pose cues for interaction anticipation.

The paper evaluates the locomotion performance and foot slip limitations of a commercial quadruped robot on sloped roof surfaces.

The paper reports field deployment findings and perception-locomotion challenges of quadrupedal robots tested in lunar regolith simulant environments.

The paper presents PLaW-VLA, a vision-language-action policy conditioning action generation on predicted latent world representations via structured attention.

The paper introduces LiteNWM, an efficient latent navigation world model that shares visual encodings to predict candidate trajectory representations.

The paper proposes a physics-informed differentiable distance learning framework using kinematics and residual networks for multi-arm collision checking.

The paper presents GLIO2, a GPU-parallel tightly-coupled LiDAR-Inertial-GNSS optimization system for robust real-time localization and mapping.

The paper introduces SCOPE, a module distilling score curvature to extract calibrated Gaussian trajectory uncertainty for real-time robot control.

The paper proposes a generative neural retargeting framework to map human manipulation demonstrations to dynamically feasible robot trajectories.

The paper proposes SpatialHarness, an online simulated scaffolding method providing complementary virtual views for fine robotic manipulation.

The paper introduces PanTact, a system for learning humanoid whole-body dexterous contact manipulation directly from wearable human demonstrations.

The paper introduces MeMoREval, a benchmark evaluating memory modalities and inference costs for embodied large language model agents.

The paper proposes graph-based heuristics combining bidirectional feasible-set generation with fixed-length searches for minimum constraint removal in motion planning.

The paper introduces an image-based 3D path-planning algorithm extending A* for navigating soft-growing robots in the spinal subarachnoid space.

The paper presents an electrothermally actuated fiber-based robot with distal segment heating for high-precision minimally invasive surgical manipulation.

The paper presents a prioritized motion-retargeting framework using an explicit priority hierarchy to transfer human motions to humanoid robots.

The paper introduces a real-robot benchmark to evaluate how reliably vision-language-action models translate spatial descriptions into physical actions.

Presents a soft pneumatic wearable support garment designed to prevent falls and assist mobility among older adults.

Presents an external revisable memory layer that tracks and retracts dynamic manipulation conditions for frozen vision-language-action policies.

Develops a simulation platform and policy framework for robot navigation combining visual observations with overheard ambient Wi-Fi signals.

Evaluates geometric verifiers that coordinate multi-robot path planning by translating trajectory conflicts into structured feedback for language models.

Presents a three-stage learning framework for whole-body reactive humanoid control from egocentric visual inputs to catch flying objects.

Combines iterative covariance steering and adaptive control to provide safe proximity navigation for spacecraft around small celestial bodies.

Presents an integrated software architecture combining stereovision, augmented reality, and tracking modules for real-time surgical robotic navigation.

Develops an online data-driven predictive control method using tactile feedback for secure and gentle robotic grasping of unfamiliar objects.

Introduces a two-agent memory architecture with an auditing mechanism to ground long-horizon episodic context into symbolic subgoals for robot policies.

Presents an automated solver for cubic spline interpolation to generate smooth, continuous waypoint trajectories for autonomous motion systems.

Proposes an action-space grounding method that projects tool affordances to transform pretrained vision-language models into robotic manipulation policies.

Proposes a counterfactual training method that mitigates head-task action bias in robot policies by isolating instruction- and action-relevant visual evidence.

Develops a temporal soft mixture-of-experts teacher-student distillation framework for versatile robot navigation across dynamic environments.

Introduces a token-allocation scheme combining sparse future references and dense execution targets within world action models.

Combines kinodynamic sampling-based motion planning with latent visual world models to solve long-horizon manipulation tasks.

Presents a lightweight vision-language-action policy trained with optical-flow future motion prediction to guide robot manipulation.

Proposes an adaptive whole-body control framework combining model predictive control, posture planning, and balance compensation for humanoids lifting unknown payloads.

Presents an imitation learning framework that jointly plans camera viewpoint repositioning and manipulation actions under partial object occlusions.

Presents a framework that recasts offline navigation demonstrations across different robot embodiments without collecting new target-embodiment data.

Proposes a guided dataset design method based on map data to develop urban perception benchmarks including heavy-duty transport modalities.

Presents a retargeting pipeline that extracts contact-aware motions from monocular human videos to teach loco-manipulation skills to humanoid robots.

Presents a replanning framework that scales trajectory revision to belief agreement for humanoid robots manipulating unfamiliar objects.

Introduces a residual action tokenization method that uses decoded reconstruction feedback to refine discrete action tokens for autoregressive robot policies.

Introduces robotic action models that supervise policy representations via action-conditioned future-state prediction while avoiding future-generation inference costs.

Introduces robotic action models that supervise policy representations via action-conditioned future-state prediction while avoiding future-generation inference costs.

Presents a wrist-centric demonstration framework that transfers human hand manipulation data to multiple distinct robotic hands without embodiment-specific demonstrations.

Enables low-latency reactive micro-aerial vehicle collision avoidance by computing Euclidean Signed Distance Field queries directly on posed depth images.

Generates physically grounded synthetic manipulation demonstrations by combining video diffusion world models with physics-engine trajectory alignment.

Accelerates constrained humanoid motion planning on GPUs by exploring tangent spaces rather than sampling ambient configuration manifolds.

Improves real-world robot reinforcement learning through selective experience reuse filtered by positive advantage estimates and human interventions.

Generates text-conditioned humanoid motions using a part-aware latent diffusion model that incorporates kinematic body structure.

Enables recursive self-improvement of endoscopic robot policies using an LLM coding agent guided by closed-loop physical experimental feedback.

Detects manipulation failures from nominal demonstrations using object-centric spatio-temporal graphs and geometric correlation scoring.

Generates and refines diverse autonomous driving trajectory hypotheses in one step using mixture-of-experts flow matching and world models.

Distills reflective diagnostic guidance into single-pass autonomous driving vision-language-action models using dual-path on-policy self-distillation.

Learns reactive humanoid tennis policies by matching generic locomotion data with internet tennis videos through trajectory-guided motion distillation.