Research
Papers, models, data and advances in robot learning.
KEMO Plugin Boosts VLA Long-Horizon Task Success by 23.6%
XDOF releases KEMO, an event-driven keyframe memory plugin for VLA models that improves long-horizon manipulation success rates by 23.6%.
Galbot Unveils AstraBrain-WBC 0.5, a Breakthrough in Humanoid Robot 'Cerebellum' Control
Galbot launches AstraBrain-WBC 0.5, the world's first general-purpose cerebellum foundation model for humanoid robots. With 80.4 million parameters, it becomes the first in its field to reach the GPT-1 scale, addressing the trade-off between agility and generalization in motion control.
Nvidia's Robot Play: Using Unitree to Build a Physical AI 'CUDA'
While Unitree faces regulatory scrutiny, NVIDIA is leveraging its Isaac GR00T reference design to build a 'CUDA-like' ecosystem for embodied AI, turning robot training into a sustainable compute business.
Embodied AI’s New Bottleneck: The Rise of a Global Data Production Chain
As Embodied AI companies face a critical data shortage, a new supply chain is emerging. This article explores how human demonstration data, particularly Ego Data from low-cost labor hubs like India, is becoming the key asset in training general-purpose humanoid robots.
Oakfruit Robotics Unveils 'Instinct-Driven' Bottom-Up Approach for Embodied AI
Oakfruit Robotics has released its 'instinct-driven' technical route for Embodied AI, adopting a bottom-up approach where robots acquire operational instincts at the foundational level and autonomously emerge with intelligent manipulation through interaction with the physical world.
X Square Robot’s WALL-OSS-0.5 Enables Zero-Shot VLA Deployment on Its Native Robot Embodiment
X Square Robot has open-sourced WALL-OSS-0.5, an Embodied AI foundation model that achieves strong zero-shot generalization on real robots directly after pre-training, challenging the industry norm that VLA models require post-training for deployment.
Behind the 2030 Revenue Target of 100 Billion Yuan: What Is AgiBot Really Aiming For?
AgiBot outlines its 'XYZ Curve' strategy, aiming for 100 billion RMB in revenue by 2030. The company is shifting from performance-based demos to practical productivity applications, introducing new models like BFM and GCFM, and pioneering a RaaS (Robotics as a Service) business model akin to AWS.
Xiaomi Open-Sources Post-Training Pipeline for Xiaomi-Robotics Embodied AI Model
Xiaomi releases the full workflow for post-training its Xiaomi-Robotics-0 embodied model on real hardware, using only 20 hours of data to master fine-grained tasks. The open-source release addresses critical challenges in action continuity and perception robustness.
AgiBot Unveils GO-2 Model to Bridge the Semantic-Motor Gap in Robotics
AgiBot released its next-generation VLA foundation model, Genie Operator-2 (GO-2), introducing 'Action Chain of Thought' and an 'Asynchronous Dual-System' architecture to address the long-standing gap between perception and execution in robotics.
Qianxun Intelligence Raises RMB 3 Billion in 30 Days, Valuing Embodied AI Firm at Over RMB 10 Billion
In just 30 days, Embodied AI company Qianxun Intelligence secured another 1 billion RMB, bringing its total financing to 3 billion RMB. With a valuation exceeding 10 billion RMB less than two and a half years after its founding, the company is accelerating its growth through a unique data-driven VLA model approach.
New Model Boosts Robot Task Success Rate to 99% and Triples Speed
Generalist AI's GEN-1 model achieves 99% success rates and 3x speed improvements by decoupling world understanding from robotic control, enabling robots to transition from demos to commercial viability.
Robotics Startup Unveils 'Fastest' World Model After Reaching a 10-Billion Valuation
StarSeaMap's Fast-WAM boosts robot reaction speed by 4x by decoupling training from inference, proving world models rely on implicit physics understanding rather than explicit future video generation.