Gemini Robotics SDK by Google DeepMind
Google DeepMind introduced a local language model for robotic control with natural-language prompts, offline functionality, and customization using 50-100 simulation demonstrations.
Key Features: - Natural language command interface - Fully offline operation - Fine-tuning with 50-100 demos - Local inference for real-time control
NVIDIA Open-Sources Isaac Sim 5.0 + Isaac Lab 2.2
Major simulation advancements include OpenUSD robot schema implementation, 10,000 FPS RGB-D synthetic data capability via MobilityGen, and Isaac Lab features for bimanual/humanoid training with multi-GPU support.
Isaac Sim 5.0 Features: - OpenUSD robot schema - 10,000 FPS synthetic data generation (RGB-D) - Photorealistic rendering - Physics-accurate simulation
Isaac Lab 2.2 Features: - Bimanual manipulation training - Humanoid robot support - Multi-GPU scaling - Streamlined workflows
Lightwheel Open-Sources SimReady Kitchen
Lightwheel released a photorealistic kitchen environment for Isaac Sim designed to support startups and researchers developing robust robotic AI.
Features: - Photorealistic kitchen scene - SimReady assets - Optimized for Isaac Sim - Free for research and development
Dex1B – 1 Billion Dexterous Hand Demos
This framework generates one billion diverse hand demonstrations for grasping and articulation tasks, enabling generative policies that transfer to real robots without fine-tuning.
Key Achievements: - 1 billion diverse demonstrations - Grasping and articulation tasks - Zero-shot transfer to real robots - No fine-tuning required
Paper: RSS 2025
DexWild (CMU) – ROS 2 Framework
RSS 2025 Best Paper winner. The framework trains manipulation policies using both robot and human demonstrations, achieving 4× improvement over robot-only models with full open-source implementation.
Key Results: - 4× performance improvement - Combines robot + human demonstrations - Full ROS 2 implementation - Open-source codebase
ROS 2 Kilted Kaiju Released
Updates include Zenoh middleware promotion to Tier 1, EventsExecutor delivering 10× Python performance gains, and simplified Windows installation.
Major Updates: - Zenoh promoted to Tier 1 middleware - 10× Python performance improvement (EventsExecutor) - Simplified Windows installation - Enhanced real-time capabilities
New ROS 2 Development Tools
rqt_tf_editor - Real-time transform visualization and editing
PlainSLAM - Minimalist LiDAR/IMU SLAM solution
lbr_fri_ros2_stack - Integration with MoveIt and Gazebo
generate_parameter_library - Streamlined configuration management
RDK X5 – ROS 2 Development Board
An AI development board (10 TOPS) with preloaded ROS 2, designed for hobbyists, students, and educators.
Specifications: - 10 TOPS AI performance - ROS 2 preloaded - Educational pricing - Maker-friendly
OpenCV Bin-Picking Challenge 2025
Top teams released their 6-DoF pose estimation code from the CVPR challenge, with validation occurring in actual robot environments rather than simulation-only approaches.