Four Gemini 4 agents beat Minecraft in a 27-hour research run
Four Gemini 4 Argon agents defeated Minecraft's Ender Dragon in one 27-hour research run. Mineflayer's game-state access and the single trial limit what the result shows.
Follow the story
spatial AI
Explore BIG CHANGE reporting on spatial AI: what changed, how it works and why it matters.
Latest reporting
20 stories
Siemens and NVIDIA test inspection and planning AI at Erlangen
Siemens says AI systems being tested at its Erlangen plant can generate inspection imagery from digital twins and propose production plans for planners to assess. The release describes live production testing, while its productivity figures remain company-reported.
EmArm uses whole-arm touch to adjust a robot's movements
HKUST's EmArm research prototype uses tactile skin and joint sensing to adjust movement. Lab tests cover touch recognition, object balancing and obstacle avoidance; deployment remains untested.
How BlenderWorld evaluates editable 3D scenes made by AI agents
QwenLM's BlenderWorld benchmark asks agents to submit native Blender projects. Its task rubrics and visual judges test different aspects of a 3D scene, with limits for real production use.
A robotic welding model tests power control before physical trials
An ISPARK preprint joins a weld-pool model, a fast surrogate and a simulated robot. Its corner test compares modelled widths; physical welding validation remains ahead.
What two-channel audio adds to Xiaomi's MiDashengLM-Spatial
Xiaomi researchers released an audio-language model for spatial questions. Its preprint reports stronger directional benchmark scores with two-channel input, while complex scenes remain difficult.
TU Delft's MonoRace paper explains its one-camera drone racing system
A new Science Robotics paper details the camera, IMU and onboard control behind a 2025 drone racing win. The result is bounded to a known competition course.
Physical Intelligence’s Pi API beta offers managed training for robot teams
Physical Intelligence’s Pi API beta lets selected robot teams upload task data, fine-tune a model on managed infrastructure and use hosted inference. Teams still need their own robots, demonstrations and repeated evaluation.
EmbodiedCLBench finds no steady gain from extra examples
EmbodiedCLBench tests simulated household tasks. One example boosts advanced-task success in the authors' study, but additional examples bring no steady gain.
OGPO trains robot policies to recover from simulated action errors
An offline training method labels progress in robot trajectories and improves simulated manipulation results under several execution disturbances. Physical-robot reliability remains untested.
H-JEPA improves visual planning in simulated tasks with a hierarchical world model
H-JEPA's three-level planner raises simulated AntMaze success from 18% to 73% with less planner compute. Its real-robot-video results measure offline path fidelity, not robot task completion.
PhoneBot puts an Android phone in charge of a small walking robot
A UCLA team's PhoneBot preprint puts sensing and walking control on an Android phone, while 13 servos and a separate onboard computer still move the prototype.
Boston Dynamics designs Atlas’s new hand for learned tool use
Boston Dynamics’ four-finger Atlas hand pairs 13 degrees of freedom, backdrivable actuation and tactile sensing with reinforcement learning in simulation. The company reports initial transfer results but no benchmark or commercial deployment.
Stuttgart funds two different robot-team approaches to construction
University of Stuttgart researchers have funding for two distinct construction robotics projects beginning in November 2026. SCALAR plans crane-assisted timber assembly; COBRAS plans mobile robot teams for reusable space frames. Neither reports site productivity results.
Japan’s humanoid data factory and national AI model program have different jobs
J-HRTI opened a humanoid data facility with 35 robots in August. Japan’s separate ¥387.3 billion FY2026 model program funds research, with no documented grant to the facility.
Runway’s Praxis-1 tests a video-trained route to robot control
Runway says Praxis-1 applies video pretraining to robot control. Its placement-error comparison offers a limited signal while public weights and fuller evaluation details remain pending.
AMD’s $8.2 billion World Labs agreement brings spatial AI research closer to chip design
AMD agreed to acquire World Labs in an all-stock deal valued at about $8.2 billion. Its spatial-model research could inform the chipmaker’s compute roadmap, but closing and product benefits remain unproven.
HomeBody connects a humanoid’s spatial memory to reusable robot skills
Caltech and Stanford researchers show a Unitree G1 using a remembered kitchen map and reusable skills. The release has selected demonstrations but no controlled results, full paper or robot code.