Verified Events
DateTypeCompanyHeadlineSegment
RLWRLD coauthors FRED world-action model with released checkpoints
- Action prediction uses future representations from one denoising step rather than iterative video generation.
- Physical Franka Research 3 experiments use a Robotiq 2F-85 gripper; OpenArm experiments test human–robot co-training.
- Released training and evaluation code accompanies 5B and 14B DROID checkpoints and a 5B RoboCasa checkpoint.
RLWRLD-coauthored VPM jointly models video and 3D point motion
- The initial VPM-Data release provides 10,000 simulation clips across tabletop manipulation and urban dynamics.
- VP-WorldBench evaluates forecasting across simulated and real settings; public implementation code accompanies the work.
- Point supervision improves the action-enabled model’s manipulation success in the authors’ simulated and DROID real-robot tests.
ACE Robotics releases ACE-Brain-0.6 spatial-grounding model
- Supervised fine-tuning uses about one million spatial-grounding examples covering roughly 940,000 images and 2.7 million question-answer pairs.
- The company reports a 74.53 average across seven public benchmarks, compared with 73.68 for GPT-6 Astra, and higher scores on six of eight evaluations.
- Its company-built RobotGrounding evaluation scored 94.13 versus 91.10; these visual-grounding results do not measure general physical-task success.
Archon and collaborators introduce PingpongDuo humanoid table-tennis doubles
- A physical clearance-and-recovery test covers 2.15 metres in2.4 seconds, with a peak speed of2.23m/s.
- Simulation comparison reports a 13.2% improvement in return success; the physical experiment measures alternating strikes.
- The system coordinates team positioning with separate footwork and striking skills.
Nature Will introduces Natura-1 dexterous physical foundation model
- The model takes language and vision inputs and controls 20-degree-of-freedom hands in a 100-millisecond loop after robot-data post-training.
- Company tests use ROKAE AR5 arms with Wuji hands and report scores of 53.4 Base and 59.1 Agentic; gripper baselines run on a different ARX platform.
- Demonstrations include long-horizon tool use, calligraphy and manipulation of deformable materials under unfamiliar instructions.
Google DeepMind broadly releases Gemini 4 Argon
- Google reports an 87% score on its Embodied Reasoning benchmark for physical and spatial problems.
- In an internal test, Argon reconstructed a human cornhole video, generated simulation code and transferred a skill to an Apptronik Apollo humanoid that made seven consecutive shots.
- The robotics developer release includes examples for simulation reconstruction, robot control, physical understanding and research reproduction.
Rhoda demonstrates tactile sensing with touch-guided Direct Video-Action models
- Tests cover normal and shear force, vibration and temperature, spanning millinewtons to over 40 newtons.
- Contact vibrations are sampled at approximately 500 Hz.
- Results are company-reported; the post does not release code, model weights or a downloadable dataset.
Tommoro Robotics introduces HABILIS Brain 0 vision-language-action model
- The company reports 95.20% success for GC-VLA and 99.55% with GCRF on the LIBERO simulation benchmark.
- Tommoro reports 400 consecutive logistics-picking successes using Rainbow Robotics RB-Y1, alongside demonstrations on UR5e and ROBOTIS AI Worker.
- The model is designed for multiple robot embodiments; manufacturing and logistics field validation remain part of the commercialization programme.
Odyssey releases Odyssey-3 foundation world model research preview
- Odyssey reports Pro scores of 66.1 on Physics-IQ Verified video-to-video and 54.7 on image-to-video using prompt enhancement and best-of-eight selection.
- The company reports first place in three of four WorldMark categories in its own evaluations.
- Robot demonstrations include recovery from missed grasps and dropped objects; physical-system adaptation uses learned action decoders or policies.
Runway introduces Praxis-1 world action model for real robots
- Noble Machines, Standard Bots and Ultra are testing the model on their hardware, including the Standard Bots RO1 arm.
- The model learns robot actions from large-scale video pretraining combined with robot interaction data.
- Public weights are planned in the coming months, following the pre-launch partner programme.