Step-by-step guides and tutorials for working with robots and robotics technology.
A practical, engineering-focused guide to reducing latency and wasted compute in VLA robot policiesâbased on StreamingVLAâs pipelined execution and FocusVLAâs task-relevant visual attention.
A practical guide to test-time policy composition (GPC) and why score-level mixing can improve diffusion and flow-based robot control without collecting new data.
A deep dive into HTD (Humanoid Transformer with Touch Dreaming) and why predicting future touch in latent space can dramatically improve contact-rich humanoid loco-manipulation.
SeedPolicy adds a self-evolving gated attention memory to diffusion policies so robots can use longer observation histories without performance collapse.
A practical guide to making robot manipulation more reliable by turning actions into monitored events, sampling efficient diffusion trajectories, and treating safety as a first class evaluation target.
A practical, end-to-end explainer of why tactile data is the bottleneck for dexterous manipulation and how new teleoperation gloves and visuotactile simulation pipelines help close the gap.
RoboPlayground reframes robot evaluation as a language-driven process over structured physical domains, letting anyone author reproducible task families that expose real generalization failures.
A practical breakdown of RGMP-S (arXiv:2601.09031): how lightweight geometric priors can ground VLM reasoning, and how spiking feature learning can improve data efficiency for robot manipulation.
A technical, practical explainer of World Action Models (WAMs) using DreamZero: how jointly modeling future video and actions improves zero-shot generalization, embodiment transfer, and real-time control.
A deep dive into OpenGo (arXiv:2604.01708), a skill-library-first architecture that uses an LLM as a dispatcher to safely switch robot-dog skills in real time.
A practical, technical guide to diffusion policies and large behavior models, why they unlocked todayâs humanoid demos, and the system-two gap that still blocks reliable deployment.
Control Barrier Function (CBF) safety filters can keep robots inside safe sets, but they can also introduce spurious equilibria or unstable behavior; this post explains why and what recent theory says to do about it.
A practical deep dive into iDP3, an improved 3D diffusion policy that drops camera calibration and segmentation by using egocentric 3D representations for humanoid manipulation.
Sensor-Conditioned Diffusion Policies show how to run robust humanoid locomotion using only onboard sensors by training a diffusion controller to predict privileged trajectories under partial observability.
A deep technical walkthrough of DexDrummer, a simulation-trained hierarchical controller that combines motion planning with residual reinforcement learning to play drums with dexterous hands.
A practical guide to policy shielding: how safety filters, constraints, and runtime monitors keep learned robot policies from doing dangerous things.
A practical look at harvest ease estimation, a decision focused approach that helps tomato harvesting robots predict success probability and adapt approach direction to improve picking reliability.
A practical guide to why touch and diffusion policies are a strong pairing for contact rich manipulation, and how recent tactile conditioned policies turn force from an accident into a controllable output.
A technical deep dive into Helix and Helix 02, Figure's pixels-to-whole-body stack that unifies perception, language, balance, and dexterous manipulation for multi-minute autonomous tasks.
LATENT shows how a humanoid can learn high-speed tennis rallies from messy, fragmentary human motion data by building a latent action space, composing primitives, and engineering robust sim-to-real transfer.
ROS-LLM connects large language model agents to the Robot Operating System so robots can turn natural language into executable behaviors with structured control, feedback, and learnable skills.
MetaWorld-X proposes a hierarchical approach to humanoid loco-manipulation that composes specialized expert policies using a VLM guided routing mechanism for more natural and stable whole body behavior.
A deep dive into the Psi0 paper and why its staged training recipe (human egocentric video pretraining plus a robot joint-space action expert) is a practical path to long-horizon humanoid loco-manipulation.
A practical deep dive into Human Action Transformer (HAT) and the PH2D dataset, and why task-oriented egocentric human demonstrations may be the fastest path to scaling humanoid manipulation.
Humanoid robots look dramatically better in demos, but reliable operation in messy human spaces still hinges on one hard problem: controlling contact forces.
OmniXtreme proposes a two stage training pipeline that scales high fidelity motion tracking across large motion libraries and still executes extreme moves on real humanoid hardware.
Learn how Isaac Lab enables GPU-parallel robot learning workflows and what to do (and avoid) when taking policies from simulation to real hardware.
A deep dive into BeetleOmni, a learning based control stack that makes an overactuated tiltable quadrotor track full 6 DoF poses with strong real world robustness.
A practical guide to diffusion policies, why they struggle out of distribution, and how online RL style fine tuning can make robot manipulation adapt during deployment.
A practical guide to the safety standards, processes, and engineering tradeoffs behind putting humanoid robots next to real people on real factory floors.
A practical guide to reducing the sim-to-real gap with domain randomization, system ID, residual learning, and modern world-model-based training workflows.
A practical deep dive into Signal Temporal Logic guided Model Predictive Control and why it helps biped robots recover from pushes, moving platforms, and unexpected terrain changes.
HAIC shows how a dynamics-aware world model can help humanoids interact with underactuated objects (like skateboards and carts) using only proprioception and a grounded dynamic occupancy map.
A new retina-inspired neuromorphic vision module detects motion first, then runs vision algorithms only where it matters, cutting latency and power for robots in dynamic environments.
A practical guide to diffusion-based robot policies, why Diffusion Transformers matter, and how on-device inference changes the economics of real-world manipulation.
A practical deep dive into why deterministic real time software matters for physical AI and how platforms like QNX are positioning themselves as the foundation for safe, reliable robots.
Sandia Labs demonstrates neuromorphic computers can solve complex physics equations with brain-like efficiency, opening new pathways for energy-efficient robotics and physical AI systems.
University of Pennsylvania researchers created the smallest fully programmable autonomous robots ever builtâsmaller than a grain of salt, powered by light, and capable of sensing, thinking, and swimming for months.
Why AI pioneers like Yann LeCun and Fei-Fei Li are betting billions on world models as the next frontier beyond language AI for robotics and autonomous systems.
Apptronik just raised $520M at a $5B valuation with backing from Google and Mercedes-Benz, bringing total Series A funding to $935M as the Austin startup races to commercialize its Apollo humanoid robots for warehouses and factories.
Algorized raised $13M to scale its edge-native Predictive Safety Engine that enables robots to predict human intent in real time, moving industrial robotics beyond detect-and-stop toward continuous, safe operation.
Alibaba launches RynnBrain, a groundbreaking AI world model that helps robots understand and interact with the physical world, marking China's ambitious entry into the physical AI race.
AGIBOT delivered a 60-minute robot-led live gala featuring dance, magic, comedy, and music performances by humanoid robots, demonstrating large-scale coordination and real-world deployment capabilities beyond industrial automation.
Carbon Robotics launched the world's first Large Plant Model, an AI system trained on 150 million plants that enables instant weed recognition without retrainingâtransforming how farmers deploy AI in the field.
Researchers from SMART and NUS have developed a brain-inspired AI control system that enables soft robots to learn tasks once and adapt in real-time without retrainingâbringing human-like intelligence to flexible machines.
Explore Physical AIâthe next frontier in robotics where AI models understand the physical world, reason about actions, and operate autonomously. Learn how NVIDIA's ecosystem is democratizing advanced robotics development.
NASA's Perseverance rover completed its first autonomous drives using vision-language models to plan routes on Mars. Learn how generative AI is transforming space exploration and autonomous navigation in the harshest environments.
Generative AI is transforming robotics by autonomously creating training data through simulation and generating new robot skills without human demonstrations. Discover how diffusion models, synthetic data generation, and self-evolving systems are accelerating robot deployment across industries.
Vision-Language-Action (VLA) models enable robots to learn manipulation tasks from human instructions and visual feedback. Discover how this breakthrough AI architecture is transforming industrial automation, research robotics, and the path to general-purpose robots.
A comprehensive guide to understanding humanoid robots: why they're the future of robotics, market size projections reaching trillions, key players from Tesla to Figure AI, and the three-phase development trajectory towards mass adoption by 2035.
Discover more robotics content across our different categories.