Robotics Research
Robo-DM Cuts Robot Dataset Storage by 70x and Decoding Time by 50x Without Accuracy Loss
Robo-DM is an open-source, cloud-based data management toolkit designed to address the storage, transfer, and training bottlenecks inherent in large-scale robot demonstration datasets. It uses EBML-based self-contained storage with both lossy and lossless compression, achieving up to 70x size reduct…
NBV Planning for Moving Object Reconstruction via Motion-Uncertainty-Aware Predictive Belief
Current next-best-view (NBV) planners optimize surface coverage but assume static objects, while motion-aware active perception methods prioritize tracking over reconstruction completeness — leaving a critical gap for dynamic scene reconstruction. This paper closes that gap with a framework that eva…
KaRMA: A Kinematic Metric That Exposes Dexterity Gaps Hidden by Static Robotic Hand Benchmarks
KaRMA (Kinematic Rolling Manipulation Ability) is a new metric designed to quantify true in-hand dexterity — the ability to continuously repose an object within a grasp — which traditional static metrics like workspace, manipulability, and grasp stability fail to capture. It models a two-finger prec…
TinySDP: Bringing Real-Time Semidefinite Programming to Microcontrollers for Certifiable Robot Motion Planning
TinySDP is the first SDP solver architected for embedded systems, enabling real-time MPC with nonconvex obstacle constraints on microcontrollers — a class of hardware previously considered too resource-constrained for SDP. The solver combines PSD cone projections with a cached-Riccati ADMM formulati…
Flow-Adaptive Ergodic Coverage Enables Formal Guarantees for Robot Exploration in Dynamic Environments
This paper reformulates adaptive robotic exploration as an ergodic coverage problem over time-varying, flow-induced domains — directly addressing the failure of static-environment assumptions in real-world settings like oceanography. The authors extend maximum mean discrepancy (MMD) as an ergodic me…
Encoder-Decoder Framework Enables Online Adaptive Dynamics for Aerial Manipulators Under Nonstationary Conditions
Aerial manipulators suffer from residual dynamics that are cross-coupled, history-dependent, and nonstationary — stemming from quadrotor-manipulator interaction, aerodynamic lag, and payload/configuration changes — which cause both physics-based and offline-learned models to degrade at deployment. T…
MuJoCo Playground: A Single-GPU Framework for Fast Robot Policy Training with Zero-Shot Sim-to-Real Transfer
MuJoCo Playground is a fully open-source robot learning framework built on MJX that integrates a physics engine, batch renderer, and training environments into a unified stack. It enables researchers to train policies in minutes on a single GPU via a simple pip install, supporting quadrupeds, humano…
Sim-to-Real Piano Playing in 30 Minutes: Residual RL Bridges the Millimeter-Precision Gap
HandelBot addresses the sim-to-real transfer failure mode for millimeter-precision dexterous tasks by layering two adaptation stages on top of a simulation-trained policy: a structured refinement pass that corrects lateral finger joint alignments from physical rollouts, followed by residual reinforc…
Atomic-Quality Probes Enable Cost-Efficient Skill-Update Governance in Compositional Robot Policies
Compositional robot skill libraries are continuously updated via fine-tuning and domain adaptation, but existing methods (BLADE, SymSkill, Generative Skill Chaining) treat the library as static at test time and cannot predict how composition outcomes change when a skill is swapped. This paper introd…
Sparse Proximity Beats Dense Tactile: Rethinking Sensor Design for Humanoid Collision Avoidance via RL
This paper presents a reinforcement learning framework for whole-body collision avoidance on the Unitree H1-2 humanoid robot, using dodgeball as a benchmark to systematically ablate distributed sensor configurations across the upper body. A key counterintuitive finding is that sparse, non-directiona…