Claire discussed Stretch robots with Aaron Edsinger from Hello Robot on the Robot Talk podcast. These robots are designed to assist people with mobility issues, helping them live more independently. Aaron Edsinger is the CEO and Co-Founder of Hello Robot, a company focused on creating practical, safe, and accessible robots. He has experience leading robotics development at Google and founded multiple companies in the robotics sector. The podcast explores the world of robotics, AI, and autonomous machines.
Intrinsic, an AI robotics group at Google, is making parts of its platform open source. Intrinsic Core™ includes capabilities like intrinsic control, pose estimation, motion planning, and grasp planning. The open source release also includes Intrinsic-ROS drivers and the Open Machine Tending Solution (OMTS) reference design. These tools aim to allow users to build sophisticated robotic applications with adaptable components. The goal is to provide developers with a starting point to customize their own physical AI solutions.
A team from Mexican EPICS in IEEE developed a portable educational platform. The project provided hands-on robotics lessons to high-school students. Participants engaged in the development and testing of RoboMeshA before its deployment. This work involved members of the EPICS in IEEE project and ITESO's Applied Professional Projects program. The project was conducted in collaboration with ITESO/Universidad Jesuita de Guadalajara.
KnowBody is a harness that makes action-relevant body relations explicit, queryable, and revisable for robot control using vision-language models (VLMs). It initializes a partial body model from a one-off trajectory to guide action selection and interpret past interactions. New evidence refines this model, and knowledge based on revised estimates is rechecked before reuse. In experiments, KnowBody achieved 75% completion on real-robot tasks compared to 25% for the native harness. Persistent updates further reduce planner rounds by 29-53% over time.
Morphometric Imitation is a three-stage framework designed to transform human hand-object interactions into zero-shot sim-to-real visuomotor policies. The first stage, morphometric optimization, kinematically retargets motions across different hand morphologies while preserving observed contacts. Subsequently, residual reinforcement learning refines these kinematic references using object pose and contact data to yield dynamically feasible robot demonstrations. Finally, these demonstrations are distilled into visuomotor policies that achieve 89.3% success in real-world trials. This method significantly improves contact fidelity and downstream retargeting success compared to baseline approaches.
OA-MPPI extends Model Predictive Path Integral control for quadrotor flight by incorporating occlusion awareness. It models regions where hidden agents might emerge from occluded areas using an extracted 3D occlusion boundary. Trajectories are penalized in the MPPI rollouts if they enter these expanding occlusion regions. This method accounts for nonlinear quadrotor dynamics and rotor thrust limits during planning. Validation in simulations and hardware experiments demonstrated improved clearance from occlusion boundaries.
This research investigates using temporal neural networks to estimate the end-effector position of an aerial continuum manipulator affected by aerodynamic disturbances. The study utilizes experimental data collected under various hovering conditions to measure positions with and without aerodynamic residuals. A strain-parameterized kinematic model forms the basis for estimating the 3D position residuals using a closed-form continuous-time (CfC) neural network. The CfC method achieved an RMSE of $22.00\pm1.70~\mathrm{mm}$, outperforming MLP and GRU methods in handling these disturbances. These results validate the effectiveness of continuous-time learning for this specific estimation task.
TapeSim is a simulator designed for efficiently modeling adhesive tape dispensing in robotic manipulation tasks. It addresses the challenge of simulating complex interactions involving flexible strips, moving rolls, and surfaces for tasks like securing wire harnesses. The simulation concentrates deformation near the unwinding region and the released strip to improve accuracy and speed. Using a rigid cluster for wound material and a deformable collar for payout simplifies interfaces and reduces simulation time. Experimental results show significant speedups and improved accuracy in simulating tape dispensing sequences.
The research introduces BirdsEye, a method for human-in-the-loop geospatial annotation using real-world field data from UAV systems. This involves recording target locations using RTK positioning and projecting these observations across multiple frames based on calibrated geometry. The method derives a mapping from camera pose uncertainty to pixel uncertainty and provides tools for sensor design based on annotation tolerances. An in-field case study demonstrated that two workers annotated significant data in approximately 12 hours. The resulting datasets allow detectors to achieve 56-89% target recovery and human review estimates detection precision at 83-87%.
OCC4M is an object-centric 4D memory designed for spatiotemporal reasoning in long-horizon manipulation tasks. It maintains persistent tracks of objects and explicitly represents temporal, motion, and containment relationships within a shared world frame. A vision-language model queries this structured memory to determine actionable targets for execution. Experimental results show OCC4M significantly outperforms a raw-history VLM baseline in memory success and end-to-end task completion. This research demonstrates the benefit of explicit, object-centric memory for complex robotic reasoning.
Streaming deep reinforcement learning is analyzed for adaptive continual learning in robotics, addressing performance degradation when robots encounter novel scenarios. The research demonstrates that streaming deep RL, following pretraining, allows robots to successfully adapt to unforeseen changes in their environment, goals, or themselves. Experiments on quadruped locomotion showed that a policy incorporating plasticity loss mitigation can adapt to diverse changes online, improving task success rates by up to 90%. Similar success was observed in robotic manipulation, though with noted stability limitations. The work concludes by discussing the limitations and future implications of this approach for continual robot learning.
FlyCNS is an embodied information-organization framework drawing inspiration from the *Drosophila* brain-nerve-cord connectome for communication-constrained control. It aims to determine which computations should remain local and which information should be transmitted for whole-body coordination. The framework preserves local sensorimotor computation while enabling selective long-range communication via separate routing pathways. It uses structural complexity from a real connectome as a weak prior for communication allocation, which guides learning alongside task-adaptive reinforcement learning. Experiments show FlyCNS achieves graceful performance degradation under tightened communication budgets, indicating the utility of neural connectome structures for guiding robot control.
KeyGen learns canonicalized semantic 3D keypoints from point clouds to create object-centric representations for policy generalization. This framework addresses the limitation of behavior cloning by focusing on object-centric features rather than instance-specific geometry and appearance. A visuomotor diffusion policy utilizes these keypoints alongside object-centric geometry to predict manipulation trajectories. Experiments demonstrate that KeyGen significantly outperforms prior methods on unseen objects and maintains robustness to pose variations and rescaling. The method achieves strong performance in both simulated and real-world manipulation tasks.