Robotics
WildCross Benchmark Enhances Robotic Perception in Natural Environments
The WildCross benchmark addresses the limitations of current robotic perception systems, particularly vision foundation models, in natural environments. It provides a large-scale dataset with synchronized RGB, depth, surface normal, Lidar, and pose data for evaluating place recognition and metric de…
OGPO: Sample-Efficient Finetuning of Generative Control Policies with Off-Policy Critics
Off-policy Generative Policy Optimization (OGPO) is a novel algorithm designed for sample-efficient finetuning of generative control policies (GCPs) in robotics. It leverages off-policy critic networks to maximize data reuse and propagates gradients through the full generative process using a modifi…
Flow Reversal Steering Enhances Generalist Robot Policy Performance Through Action Space Steering
Flow Reversal Steering (FRS) is a novel method designed to enhance the performance of generalist robot policies, particularly in complex tasks. By reversing the flow of a policy, FRS can infer and invoke appropriate actions from a policy's latent space, even when direct commands are suboptimal. This…
Heterogeneous LiDAR Early Fusion with Learned Re-Ranking Boosts Agricultural Place Recognition by 30%
MinkUNeXt-VINE++ tackles long-term place recognition in unstructured agricultural environments by fusing heterogeneous LiDAR point clouds from two distinct sensors (Livox Mid-360 and Velodyne VLP-16) at the input level and applying a learned re-ranking step at inference to disambiguate repetitive sc…
Mana: Animating Dexterous Manipulation for Articulated Tools
Mana is a sim-to-real framework that tackles articulated tool manipulation by reframing it as an animation problem. It utilizes a coarse-to-fine pipeline, combining motion planning and reinforcement learning to generate manipulation trajectories from grasp keyframes. This approach significantly auto…
Humanoid robots learn self-other distinction from proprioceptive-visual correspondence without labels or models
This paper demonstrates that a humanoid robot can learn to distinguish itself from other agents (including morphologically identical robots and humans) by leveraging the correspondence between its own proprioceptive signals and visual observations, requiring no identity labels or kinematic priors. T…
StereoPolicy: Enhancing Robotic Manipulation with Implicit Geometric Reasoning via Stereo Vision
StereoPolicy is a novel framework designed to improve robotic manipulation by directly utilizing synchronized stereo image pairs. It employs a Stereo Transformer with cross-attention to implicitly capture geometric and disparity cues, enhancing reasoning without explicit 3D models. This approach has…
One VLA Policy for All Robots: Decentralized Multi-Embodiment Coordination Without Communication
CHORUS shows that a single pretrained vision-language-action (VLA) model backbone can be reused across heterogeneous robots to enable decentralized multi-robot collaboration, requiring only local observations and a robot-identifying prompt at inference. By leveraging visuomotor priors from VLA pretr…
FACTR 2: Force Estimation and Enhanced Policy Learning for Commodity Robots
This work introduces Neural External Torque Estimation (NEXT), a data-driven approach for estimating external joint torques on robot arms without requiring dedicated force sensors. NEXT achieves comparable accuracy to traditional sensors with minimal training data and time. It enhances policy learni…
Integrated Tactile Sensor with Deformable Pad Enables Slip-Aware In-Hand Manipulation
This paper presents a novel, deformable tactile sensor for in-hand manipulation, uniquely integrating velocity, force/torque, and pressure map sensing within a single compliant structure. Designed for rapid, low-cost fabrication, the sensor demonstrates robust tracking of diverse surfaces and is int…
DuoBench: A New Benchmark for Evaluating and Improving Bimanual Robot Manipulation
DuoBench is introduced as a novel, extensible benchmarking framework for evaluating bimanual manipulation policies on the FR3 Duo platform, addressing the limitations of existing benchmarks in capturing dual-arm control complexities. It features eleven tasks across four coordination categories, impl…
Sim-to-Real Tactile VLA via Co-Training and Reinforcement Learning
TacCoRL injects tactile feedback into vision-language-action (VLA) policies without requiring large-scale tactile pretraining or extensive real-world contact exploration. The framework uses a real-aligned simulator for closed-loop contact training, where mixed simulated and real trajectories warm-st…
SARM2: Multi-Task Reward Modeling Enhances Robotic Manipulation Policy Improvement
This work introduces RM, a multi-task, stage-aware reward model that improves value estimation accuracy by 80% for robotic manipulation tasks. RM integrates an action-primitive-based stage estimator with an MMoE value head. The SPIRAL framework leverages RM for self-improvement of VLA policies using…
Three-Stage Motion-Guided Curriculum RL Enables Humanoid Soccer Shooting with Near-Professional Velocity
RoboNaldo introduces a three-stage curriculum reinforcement learning framework that uses a single human kick reference as a scaffold, progressively shifting optimization from whole-body stability to shooting performance. This overcomes the core tension between motion tracking (stable but inflexible)…
Embodiment-Conditioned Generalist Control Policy for Diverse Multirotor Aerial Robots
This work introduces a generalist position control policy for multirotor aerial robots, capable of controlling diverse configurations (e.g., quadrotors, hexarotors) with a single neural network. The policy leverages a physics-grounded embodiment descriptor, a normalized control allocation matrix, to…
MemoryVLA++: Enhancing Robotic Manipulation with Temporal Modeling via Memory and Imagination
MemoryVLA++ introduces a novel temporal modeling framework for Vision-Language-Action (VLA) models in robotic manipulation. It enhances performance on long-horizon tasks by integrating working memory, episodic memory retrieval from a Perceptual-Cognitive Memory Bank, and future state imagination via…
Biaxial-Tilt Hexacopter Maintains Full 6-DOF Control Through Multiple Simultaneous Rotor Failures Without Fault Detection
This paper demonstrates that a biaxial-tilt overactuated hexacopter (BTO) can passively maintain full 6-degree-of-freedom actuation following abrupt, unknown total rotor failures — a regime where conventional multirotors lose controllability entirely. Two fault-tolerant schemes are proposed: one at …
Center-of-Pressure Tactile Representation Enables Zero-Shot Sim-to-Real Transfer for Dexterous Manipulation
The sim-to-real gap for tactile sensing in dexterous manipulation has forced prior methods to collapse rich touch data into low-dimensional binary features, discarding the very information needed for complex tasks. This paper introduces Center-of-Pressure (CoP), a physics-grounded tactile representa…
QDTraj: Enhancing Robotic Manipulation through Diverse Trajectory Primitive Generation
QDTraj utilizes Quality-Diversity algorithms with sparse reward exploration to generate a diverse set of high-performing trajectory primitives for robotic manipulation. This approach enables robots to learn multiple solutions for tasks, improving adaptability to real-world constraints and unexpected…
Dexterous Tendon-Driven Wrist Achieves 99% Handkerchief Unfolding via Hierarchical Control and Particle-Spring Modeling
Researchers developed a parallel anti-parallelogram tendon-driven wrist enabling 90-degree omnidirectional rotation with low inertia and decoupled roll-pitch sensing for spinning flexible objects like handkerchiefs. A high-low hierarchical control scheme paired with a particle-spring model abstracts…
Switch: Hierarchical Multi-Skill System for Agile Humanoid Locomotion
The "Switch" system addresses limitations in humanoid robot skill transitions by introducing a hierarchical multi-skill framework. This framework utilizes a Skill Graph (SG) for kinematically similar transitions, a deep reinforcement learning-trained whole-body tracking policy, and an online skill s…
Benchmarking Classical Coverage Path Planning on Irregular Hexagonal Grids
This paper introduces a reproducible benchmark for deterministic single-vehicle coverage path planning heuristics on irregular hexagonal grids, relevant for maritime applications. It addresses the lack of standardized comparisons for these methods, which are often evaluated on limited examples or in…
D-REX: Differentiable Real-to-Sim-to-Real Engine for Dexterous Grasping
D-REX introduces a differentiable real-to-sim-to-real engine leveraging Gaussian Splat representations for robotic systems. This engine aims to bridge the simulation-to-real-world gap by enabling object mass identification from visual observations and control signals, while simultaneously facilitati…
Robotic Peeling with Human Preference Alignment
This paper presents a two-stage learning framework for fine-grained robotic manipulation tasks with subjective success criteria, using knife-peeling as a case study. The approach combines force-aware imitation learning for robust initial policy generation with preference-based finetuning using a lea…
XL-VLA: A Cross-Embodiment Latent Space for Dexterous Robot Manipulation
XL-VLA introduces a novel vision-language-action framework that utilizes a unified, embodiment-invariant latent action space. This approach enables scalable cross-embodiment training and efficient data reuse for dexterous manipulation tasks, addressing the challenge of costly data collection for div…
Physics-Aligned Simulation for Deformable Object Manipulation
Robotic manipulation of deformable objects is data-intensive due to complex shape, contact, and topological changes. Traditional sim-to-real methods fail because simulators are ungrounded and lack accurate physics for these objects. SIM1, a physics-aligned data engine, addresses this by digitizing s…
AEROS: A Single-Agent Operating System for Embodied AI
AEROS presents a novel architectural paradigm for robotic systems, treating a robot as a single persistent intelligent agent. Its core innovation lies in the use of Embodied Capability Modules (ECMs) that encapsulate skills, models, and tools, enforced by a policy-separated runtime. This design enha…
RoSHI: A Hybrid Wearable for Robust Human Pose and Shape Estimation in Robotic Learning
RoSHI is a novel hybrid wearable system designed to capture rich, long-horizon human interaction data for scaling robot learning. It integrates low-cost sparse IMUs with Project Aria glasses to achieve precise 3D pose and body shape estimation in a global coordinate frame. This approach addresses li…
EgoVerse: Scaling Robot Learning Through Egocentric Human Data, Bypassing Teleoperation
EgoVerse leverages egocentric human data to scale robot learning, moving beyond traditional teleoperation. This approach, supported by the EgoScale and dexterity scaling law, uses behavior cloning from human actions to enhance robot capabilities without direct robot interaction during the learning p…
Robotics Software Lags Hardware, Hampered by Reliability and Misaligned AI
Robotics development is currently bottlenecked by hardware reliability issues, which slow down software iteration despite advanced physical capabilities. The field also suffers from a lack of standardized benchmarking, leading to irreproducible results and difficulty in objective comparison. Further…
Amazon FAR Open-Sources Holosoma: A Full-Stack Robotics Platform for Sim-to-Real Transfer
Amazon FAR has open-sourced Holosoma, a comprehensive robotics platform designed to address the full-stack challenges of sim-to-real learning for humanoid robots. Holosoma provides a unified framework supporting multiple simulation backends (IsaacGym, IsaacSim, MJWarp), various robots (humanoid and …
Robotics: Conquering the Physical World with AI
While AI has largely conquered the digital domain, the next grand challenge lies in mastering the physical world. This requires a data maximalist and model minimalist approach, leveraging synthetic data generated through advanced simulation and video world models. The ultimate goal is to achieve a "…
World-Model-Guided Trajectory Generation Unlocks One-Shot Robot Imitation on Unseen Tasks
OSVI-WM addresses a critical gap in one-shot visual imitation learning: generalizing to unseen tasks that are visually similar to training tasks but require semantically distinct responses. The framework uses a learned world model to predict latent state-action trajectories from a single expert vide…
RoboCulture Enables Cost-Effective Robotic Automation of Long-Duration Biological Experiments
RoboCulture is a flexible, low-cost platform using a general-purpose robotic manipulator to automate biological workflows, addressing limitations of current liquid handlers that require human intervention for plate loading, tip replacement, and calibration. It integrates liquid handling, lab equipme…