Back to Newsletter Archive
Issue #4
July 2, 2025

Gemini Robotics SDK, Isaac Sim 5.0, Dex1B

AI SDKSimulationDexterous ManipulationROS 2

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.

Enjoyed this newsletter?

Get weekly robotics updates delivered to your inbox.

Subscribe to Newsletter