Technical Directions
Robotics coverage relating to robot policies.
142 selected reportsLast 30 days 100 recordsAll coverage 811 records
Skild AI has acquired Zebra Technologies' robotics division to deploy its omni-bodied brain across warehouses, aiming to unlock massive productivity gains and enhance end-to-end fulfillment processes. This acquisition integrates Zebra's Symmetry Fulfillment platform with Skild's foundation model, enabling cross-embodiment generalization and intelligent decision-making in logistics environments.
Dexterity has deployed Instinct, a tactile intelligence system that enables robots to perform force-guided manipulation with zero retraining. Instinct reacts to contact forces in 0.5–2 ms, 10–100× faster than human touch, and adapts in real time to constrained placements. It integrates with Dexterity’s world model, Foresight, to update contextual understanding through contact data, achieving an 81% success rate in production with adaptive strategies.
Google DeepMind has released Gemini Robotics-ER 1.6, an upgraded model that enhances embodied reasoning for robotics. This model improves spatial reasoning, multi-view understanding, and safety compliance, enabling robots to perform complex tasks like instrument reading. It is available via the Gemini API and Google AI Studio, with examples provided in a developer Colab.
Helix 02, a single neural system that controls the full body directly from pixels, now performs end-to-end living room cleanup, including tasks like cleaning surfaces, handling flexible objects, and bimanual manipulation. The system learns new capabilities through data without requiring new algorithms or specialized engineering.
This tutorial explores the challenges of deploying VLA models on embedded robotic systems, including dataset recording best practices, fine-tuning techniques for ACT and SmolVLA, and real-time performance optimization using the NXP i.MX 95 SoC. It emphasizes asynchronous inference and hardware-aware scheduling to improve control and reduce latency.
Dexterity introduces Foresight, a world model that enables robots to reason about the physical world, predict outcomes, and act confidently in real environments. Trained on over 100 million autonomous actions in production, Foresight is designed for interpretability, safety, and performance, with capabilities like predictive branching and pragmatic decision-making. It addresses complex tasks like spatial packing and includes a game and API challenge for further exploration.
Physical AI refers to AI systems that enable robots to perform physical tasks in the real world. The article explains that this is extremely challenging due to the infinite variability of the physical environment, requiring solutions across perception, planning, motion control, and safety. Dexterity's compositional approach, using specialized AI agents and an orchestrator, enables production-scale Physical AI with over 100 million autonomous actions executed in real enterprise operations.
LeRobot introduces a Vision-Language-Action (VLA) policy called PI0Fast, which predicts continuous robot actions via autoregressive next-token prediction over FAST action tokens. The repository includes training and inference code, along with instructions for evaluation in simulation and real-world settings.
Dexterity introduces 'Transactable World Models' as a foundational component for Physical AI, emphasizing interpretability, physics integration, and real-time consistency. The approach treats world models as operators rather than data stores, enabling multi-agent coordination and robust manipulation in complex environments. This contrasts with traditional world models that focus on data compression and replay, lacking causal reasoning and physical grounding.
This hands-on tutorial walks through the process of collecting data, training policies, and deploying autonomous medical robotics workflows on real hardware using NVIDIA Isaac for Healthcare. It introduces the SO-ARM starter workflow, which enables developers to build and validate surgical assistant robots from simulation to deployment.
Hugging Face announces LeRobot v0.4.0, a major upgrade for open-source robotics with Dataset v3.0, new VLA models like PI0.5 and GR00T N1.5, and a plugin system for hardware integration. The release also adds support for LIBERO and Meta-World simulations, multi-GPU training, and a new Hugging Face Robot Learning Course.
Dexterity's Mech industrial superhumanoid has been approved for onsite operational validation at Sagawa Express's X Frontier® relay center in Tokyo, Japan. This marks the start of Mech's use at Sagawa, with plans to scale the technology to additional relay centers. The deployment aims to address labor shortages and regulatory challenges in Japan's logistics industry.
LeRobot has released a Vision-Language-Action (VLA) model, π₀.₅, trained on robot demonstrations and multimodal data for long-horizon tasks. The model supports flow-matching for action prediction and is compatible with the original reference code. It is intended for fine-tuning on specific use cases and includes both inference and training examples.
LeRobot has released π₀.₅, a Vision-Language-Action (VLA) model with open-world generalization capabilities, trained on robot demonstrations and multimodal data. The model supports flow-matching for action prediction and is intended as a base model for fine-tuning. Code and documentation are provided for inference and training.
LeRobot introduces π₀, a Vision-Language-Action (VLA) foundation model that jointly reasons over vision, language, and actions to control robots. The model is designed as a base architecture for enabling open-world generalization and includes training and inference instructions for deployment.
LeRobot has released the π₀.₅ (Pi05) model, a Vision-Language-Action (VLA) model with open-world generalization capabilities. It is trained on robot demonstrations and multimodal data to execute long-horizon tasks in real-world environments. The model supports flow-matching for action prediction and is intended as a base for fine-tuning on specific applications. The implementation is compatible with the original reference code and includes instructions for inference and training.
LeRobot releases the π₀ model, a Vision-Language-Action (VLA) foundation model that jointly reasons over vision, language, and actions to control robots. It serves as the base architecture for π₀.₅'s open-world generalization. The model supports training, inference, and real-world evaluation with detailed code examples and setup instructions.
NVIDIA has released the GR00T N1.5 model, a cross-embodiment foundation model for generalized humanoid robot reasoning and skills. The model can be fine-tuned using teleoperation data from a SO-101 arm, with a detailed tutorial provided for developers. The release includes instructions for dataset preparation, fine-tuning, evaluation, and deployment.
Hugging Face introduces SmolVLA, a 450M parameter open-source Vision-Language-Action model for robotics, trained on community-shared datasets. It outperforms larger models in simulation and real-world tasks, supports asynchronous inference for faster response, and is designed for deployment on consumer hardware.
SmolVLA is a compact, efficient Vision-Language-Action (VLA) model designed for affordable robotics, trainable on a single GPU and deployable on consumer hardware. It matches the performance of larger VLAs through community-driven data, with a reference implementation and training instructions provided.
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).