🤖 Amazon Gives Home Robots Another Shot

Good Morning, Roboticists!
Amazon wants another robot in your house. At this point, Prime might eventually include an actual employee living with you.
Amazon Gives Home Robots Another Shot

TL;DR: Amazon is assembling a new home robotics effort after Astro’s limited rollout, according to job listings analyzed by CNBC. The Personal Robotics Group combines redeployed Astro staff with Fauna Robotics, which Amazon acquired in March. Recruitment includes customer support, repairs and interaction design. The company has disclosed no final robot design, price, feature list or launch date. Read more →
One Policy Grabs Sixty-Six Objects

TL;DR: Amazon FAR’s Dex-Hands team demonstrated one policy grasping 66 objects, from AirPods to dumbbells and pipe wrenches. Jitendra Malik says it learned from video and reinforcement learning, with no teleoperation data. The system uses head-mounted stereo vision and Sharpa Wave hands. Malik clarified that similar objects appeared in training; a formal success-rate benchmark has not been released. Read more →
A $320 Gripper Trains Robots

TL;DR: Berkeley researchers released Specialist’s gripper build files and manuals alongside an open robot-learning pipeline. The $320 handheld device records demonstrations using an Insta360 camera and printed tags. A matching actuated gripper lets learned policies run without retargeting. The team reports three-millimeter median tracking error across eleven sessions, while noting roughly 150 milliseconds of camera delay during deployment. Read more →
Thirty Kilos Ride Behind Two Hands

TL;DR: A Unitree G1 pushed a cart carrying up to 30 kilograms at 0.5 meters per second using ForPC, a controller trained entirely in simulation. It infers interaction forces from onboard signals rather than external force sensors. Physical tests included narrow passages, turns and a payload change at a transfer station, without retuning the policy between transport segments. Read more →
The Arm Takes A Detour

TL;DR: Munich researchers added obstacle guidance to Franka robot policies without collecting new demonstrations or retraining them. Across seven physical scenarios, the system completed 105 of 140 tasks without collisions, including pouring, pen insertion and putting a sock into a drawer. Obstacles were measured manually, and guidance strength was tuned separately for each scenario before the repeated trials. Read more →
Smooth Moves Beat Action Chunks

TL;DR: ODEAction replaces separately predicted action chunks with continuously evolving control commands. On an AgileX Cobot Magic dual-arm robot, researchers tested banana placement, drawer opening, cup placement and spoon insertion. Across twenty trials per task, average success rose from 65% to 80% against the matched base policy. Cup placement required carrying water without spilling throughout the motion. Read more →
The Lab Camera Chooses Its View

TL;DR: Empa and ETH Zurich researchers trained a robotic camera to find and frame hydrofoil wakes in a water tunnel. Learning happened entirely in simulation before transfer to the physical arm. The policy resolved flow details with fewer images and roughly half the time of a full scan. The experimental setup cost about €15,000; exposure and aperture remained fixed. Read more →
A Soft Cup Measures The Squeeze

TL;DR: Italian and Dutch researchers tested a soft suction cup that uses optical signals to estimate contact forces on kiwifruit. On fruits excluded from training, the best model’s mean absolute error was 0.138 newtons. These were controlled compression tests without active suction. Firmness classification was less reliable, and integrating contact detection with suction grasping remains future work for the team. Read more →
The Rover Checks Its Own Limits

TL;DR: Tartu researchers taught a Jackal mobile robot to judge routes by predicted execution rather than terrain geometry alone. The model learned from simulated Husky runs before transfer to the physical platform. It considers whether the robot can track a route and keep moving. The current implementation stops to replan, taking about one second, and assumes firm terrain. Read more →
AgiBot Finds A Qatar Network Partner

TL;DR: Vodafone Qatar and AgiBot signed an agreement at MWC Doha to explore embodied robotics for government, businesses and education. Vodafone will assess how its connectivity and digital infrastructure could support robot operation, while the partners identify sector-specific applications. The announcement names no customer deployment, purchase commitment, robot count or delivery schedule, making this a market-entry partnership rather than an order. Read more →
TOGETHER WITH US
AI Secret Media Group is the world’s #1 AI & Tech Newsletter Group, reaching over 2 million leaders across the global innovation ecosystem, from OpenAI, Anthropic, Google, and Microsoft to top AI labs, VCs, and fast-growing startups.
We've helped promote over 500 Tech Brands. Will yours be the next?
Email our co-founder Mark directly at mark@aisecret.us if the button fails.