Fundamentals
From LLMs to robots that act: understand the VLA revolution, from essential concepts to the safety standards that govern human-robot collaboration.
Learning Path Map
🌍 The Big Picture
From LLMs to autonomous robots: the complete evolution
🔬 Anatomy of a VLA
Vision + language + action: breaking down the architecture
🗺️ Model Overview
RT-2, OpenVLA, pi0, Qwen-VLA, and 8 more models
📦 Datasets
1M+ trajectories: Open X-Embodiment and DROID
🤖 Hardware
Cobots, humanoids, and accessible platforms
🛡️ Safety and Standards
ISO 10218:2025 and safety-by-design for VLAs
Detailed content
From LLMs to Robots That Act: The Big Picture
The evolution of AI: from text models to agents that manipulate the physical world.
1 From language to action: LLM → VLM → VLA
What it is
The progression from models that process only text (LLMs) to models that also see (VLMs), and finally to models that generate motor actions in the real world (VLAs). Each leap adds a modality: first vision as input, then motor control as output.
Why learn
Understanding this progression is the foundation for everything that comes later in the course. Without clarity on what changed between GPT-4V and RT-2, it is impossible to evaluate VLA architectures or decide which model to use in a real project.
Key concepts
An LLM (Large Language Model) processes sequences of text tokens. A VLM (Vision-Language Model) adds a visual encoder to process images alongside text. A VLA (Vision-Language-Action) extends the VLM with an action head that decodes tokens into continuous motor commands such as position (x,y,z) and gripper rotation.
2 The gap between understanding and acting
What it is
The fundamental gap between semantic understanding (knowing what a cup is) and motor competence (being able to pick up the cup without dropping it). VLMs understand visual scenes accurately, but cannot translate that understanding into coordinated physical actions.
Why learn
This gap defines the central problem VLAs try to solve. Recognizing it helps you understand why simply connecting ChatGPT to a robotic arm doesn't work, and which training techniques are needed to close this gap.
Key concepts
Grounding is the process of anchoring abstract concepts in concrete sensory perceptions. The sim-to-real gap refers to the difference between performance in simulation and in the real world. Closed-loop control is the system's ability to continuously adjust actions based on sensory feedback.
3 Foundation models for robotics
What it is
Applying the foundation model paradigm (large-scale pretraining + fine-tuning) to robotics. Instead of training a model from scratch for each task and robot, a base model trained on massive datasets from multiple robots and environments is adapted.
Why learn
Foundation models are why VLAs became viable. Pretraining on diverse internet data (vision, language) provides rich representations that transfer surprisingly well to robotic control, even with few demonstration data points.
Key concepts
Transfer learning lets you reuse knowledge from data-rich domains (the internet) in data-scarce domains (robotics). Multimodal pretraining combines vision and language in a unified representation. Cross-embodiment generalization lets a single model control robots with different morphologies.
4 Affordances: what the robot CAN do
What it is
A concept from J.J. Gibson's ecological psychology adapted for robotics: affordances are the action possibilities that an object or environment offers a specific agent. A cup “affords” being grasped; a door “affords” being opened. For a VLA, detecting affordances means identifying which actions are physically feasible in the current scene.
Why learn
VLAs need not only to understand natural language instructions, but also to assess whether the requested action is possible given the current state of the world. Affordance detection is what distinguishes a robot that tries blindly from one that plans feasible actions.
Key concepts
Visual affordances map image regions to possible actions (graspable surfaces, navigable areas). Google's SayCan combines an LLM that proposes actions with a value function that filters for physically feasible actions. Task and Motion Planning (TAMP) integrates symbolic planning with continuous motion planning.
5 Embodied AI vs. disembodied AI
What it is
The distinction between AI that exists only in digital space (chatbots, image generators) and AI embodied in physical bodies that interact with the real world. Embodied AI faces challenges absent from traditional AI: Newtonian physics, sensory latency, mechanical wear, and irreversible consequences of errors.
Why learn
Working with VLAs requires a shift in mindset compared with conventional AI. Errors aren’t just wrong answers in a chat; they’re collisions, dropped objects, or safety risks. Understanding this fundamental difference is a prerequisite for designing responsible robotic systems.
Key concepts
The embodiment hypothesis holds that true intelligence requires a physical body that interacts with the environment. Sensorimotor contingencies are the regularities between actions and sensory changes that an embodied agent learns. The reality gap is the mismatch between simulation and the real world that makes deploying embodied AI much more difficult.
6 Timeline: from GPT to autonomous humanoids
What it is
The timeline of milestones from the first modern LLMs (GPT-3, 2020) to today’s autonomous humanoids (Figure 02, GR00T, 2025-2026). It includes the release of RT-2 (2023), the founding of Physical Intelligence (2024), OpenVLA (2024), and the explosion of robotics startups in 2025.
Why learn
The pace of progress is crucial for calibrating expectations. In just 3 years, we went from "LLMs can't control robots" to "generalist VLAs control multiple embodiments." Knowing the timeline helps you project where we'll be in 2027-2028.
Key concepts
Scaling laws for robotics investigate whether more data and parameters improve motor performance the way they improve language performance. Emergent capabilities in VLAs refer to skills that arise with scale without being explicitly trained. The open-source ecosystem (LeRobot, OpenVLA, Octo) democratizes access that was previously limited to labs with millions in hardware.
Anatomy of a VLA: Vision + Language + Action
Breaking down the architecture: visual encoder, language backbone, and action decoder.
1 Visual encoder: SigLIP, DINOv2, ViT
What it is
The component that converts camera images into vector representations (visual tokens) that the LLM backbone can process. SigLIP uses contrastive learning with sigmoid loss, DINOv2 uses self-supervised learning without labels, and ViT (Vision Transformer) divides the image into 14x14 or 16x16 pixel patches.
Why learn
The quality of the visual encoder determines what the VLA "sees." A weak encoder misses crucial details such as object edges or textures that indicate fragility. The choice between SigLIP and DINOv2 directly affects performance on fine manipulation tasks.
Key concepts
Patch embeddings divide the image into blocks and project each one into a dense vector. Contrastive learning (SigLIP) aligns images and text in the same vector space. Self-supervised features (DINOv2) capture geometric structures without human supervision, making them especially good for estimating depth and segmenting objects.
2 LLM backbone: Llama, Gemma, Qwen
What it is
The VLA’s central brain: a pretrained LLM that fuses visual tokens with text tokens in a unified representation space. Models such as Llama 3 (Meta), Gemma 2 (Google), and Qwen 2.5 (Alibaba) serve as backbones because they already have reasoning capabilities and world knowledge from billions of text tokens.
Why learn
The LLM backbone gives the VLA the ability to understand complex natural language instructions and reason about scenes. The choice of backbone defines the trade-off between capacity (7B, 13B, 72B parameters) and inference latency, which is critical for real-time control.
Key concepts
The projection layer maps visual tokens to the LLM embedding space (typically a simple MLP). Instruction tuning adapts the backbone to follow robotic commands. LoRA (Low-Rank Adaptation) enables efficient fine-tuning of the backbone without retraining all 7B+ parameters.
3 Action head: tokenization of continuous actions
What it is
The final component that converts the LLM’s internal representation into executable motor commands. A robot’s actions are continuous (position x=0.342m, torque=1.7Nm), but LLMs operate with discrete tokens. The action head resolves this incompatibility through discretization or generative models.
Why learn
The action head is where VLA diverges radically from a standard VLM. It is the architectural innovation that makes it possible to generate physical actions. Different designs (MLP, diffusion, flow matching) produce qualitatively different behavior in terms of smoothness, multimodality, and precision.
Key concepts
Discretization bins divide the continuous action space into N bins (typically 256) and treat each action as a categorical token. A diffusion action head uses an iterative denoising process to generate smooth, multimodal actions. Flow matching (used in pi0) learns a vector field that transforms noise into actions via an ODE.
4 Trajectory Tokenization: Discrete vs. Continuous Bins
What it is
Two fundamental approaches to representing robotic trajectories as processable sequences. In bin discretization, each action dimension (x, y, z, rotations, gripper) is quantized into buckets. In the continuous approach, the model directly generates real-valued vectors through regression or generative models.
Why learn
This choice directly affects control accuracy. Discrete bins are simple and compatible with the LLM vocabulary, but lose resolution. Continuous approaches are more accurate, but require specialized architectures. Understanding the trade-off is essential for evaluating VLA models.
Key concepts
Action tokenization with 256 bins per dimension provides ~0.4mm resolution in a 10cm workspace, sufficient for most tasks. Action chunking predicts multiple future actions at once (typically 4-16 steps), improving temporal consistency. De-tokenization converts the predicted tokens back into continuous commands for the robot controller.
5 Co-fine-tuning: vision + language together
What it is
The process of training the visual and language components of a VLA simultaneously on robotic demonstration data. Instead of freezing the visual encoder and training only the action head, co-fine-tuning updates all components together so they learn representations optimized for motor control.
Why learn
Co-fine-tuning is what enables VLAs to outperform modular pipelines (perception separated from control). OpenVLA demonstrated that training vision and language together improves performance by up to 16.5% compared with fine-tuning only the action head, justifying the extra computational cost.
Key concepts
End-to-end training propagates gradients from the action loss back to the visual encoder, adapting visual features to the task. Catastrophic forgetting is the risk of losing original language capabilities during fine-tuning, mitigated by regularization and a low learning rate. Multi-task fine-tuning trains on multiple tasks simultaneously to maintain generality.
6 Latency and throughput: token budget
What it is
The real-time constraints VLAs face: a robot needs to choose actions at 5-50 Hz, but a 7B-parameter LLM can take hundreds of milliseconds per token. The token budget determines how many visual tokens and action tokens the system can process within the available control time.
Why learn
Latency is the most critical practical bottleneck in real VLAs. A brilliant model that takes 500ms per action is useless for dynamic manipulation. Understanding the token budget helps you choose between larger models (better, but slower) and compact models (faster, but limited).
Key concepts
Control frequency determines how many times per second the model needs to generate actions (5 Hz for slow tasks, 50 Hz for dexterous manipulation). KV-cache lets you reuse computations from previous tokens, speeding up inference. Quantization (INT8, INT4) reduces model size by 2-4x with minimal performance loss, enabling deployment on embedded GPUs such as Jetson Orin.
Model Overview: RT-2, OpenVLA, π0, Qwen-VLA
The VLA model ecosystem, from closed pioneers to open source.
1 RT-2 and RT-X: Google DeepMind's pioneers
What it is
RT-2 (Robotics Transformer 2) was the first VLA to co-train robotic data with large-scale web data, demonstrating that knowledge from the internet transfers to motor control. RT-X extended this idea by training on the Open X-Embodiment dataset with multiple robots, creating a generalist cross-embodiment model.
Why learn
RT-2 defined the paradigm that virtually all VLAs follow today. Understanding its design choices and limitations (closed model, 55B parameters, slow inference) provides context for why the community sought open and efficient alternatives like OpenVLA and Octo.
Key concepts
Web-scale knowledge transfer reuses the semantic knowledge of VLMs trained on internet data to recognize objects never seen in demonstrations. Co-fine-tuning mixes batches of web and robotics data in the same training run. RT-2 represents actions as text tokens, reusing the LLM’s vocabulary to issue commands.
2 OpenVLA: Stanford’s open-source standard
What it is
OpenVLA is a fully open-source 7B-parameter VLA model, trained on 970 thousand Open X-Embodiment trajectories. It combines two vision encoders (SigLIP and DINOv2) with the Llama 2 backbone, achieving performance competitive with RT-2-X using a fraction of the parameters.
Why learn
OpenVLA is probably the model you’ll use in practice to experiment and fine-tune. Because it’s open, with weights and code available, it has become the reference baseline for academic research and the natural starting point for real projects with limited budgets.
Key concepts
The dual visual encoder (SigLIP + DINOv2) combines semantic and geometric features, improving fine manipulation. Llama 2 7B serves as the reasoning backbone. Efficient fine-tuning via LoRA runs on a single consumer GPU, and quantization enables inference at ~6 Hz.
3 π0 (pi0): flow matching from Physical Intelligence
What it is
π0 is Physical Intelligence’s flagship VLA, built on the PaliGemma backbone and using flow matching to generate high-frequency continuous actions (up to 50 Hz). It’s designed for complex dexterous manipulation, such as folding clothes and assembling boxes, across multiple embodiments.
Why learn
π0 represents the state of the art in smoothness and control frequency, areas where models based on discrete tokenization fall short. Studying it shows how modern generative techniques (flow matching) unlock fine manipulation that was previously impractical with VLAs.
Key concepts
Flow matching learns a vector field that transports Gaussian noise to the action distribution via an ODE, generating smooth trajectories. Action experts are subnetworks dedicated to action generation, coupled to the VLM backbone. Control at 50 Hz enables high-dexterity tasks with nearly continuous feedback.
4 Octo: Berkeley's generalist transformer
What it is
Octo is a compact generalist model (27M to 93M parameters) based on a transformer, trained on 800 thousand trajectories from Open X-Embodiment. Unlike giant VLAs, it was designed to be modular: new sensors and action heads can be added without retraining everything.
Why learn
Octo shows that not every VLA needs billions of parameters. Its flexible, lightweight architecture is ideal for rapid research and deployment on modest hardware, making it a practical alternative when OpenVLA is too large for the compute budget.
Key concepts
The architecture uses a transformer with attention blocks over observation and task tokens. Modular design makes it possible to plug in new input heads (cameras, proprioception) and output heads without rewriting the core. A diffusion action head generates smooth multimodal actions, and the model supports conditioning by language or by goal image.
5 Qwen-VLA, SmolVLA, and compact models
What it is
A new generation of compact VLAs optimized to run on low-cost embedded hardware. SmolVLA (from Hugging Face/LeRobot) has fewer than 500M parameters and runs on a CPU; Qwen-VLA takes advantage of Alibaba’s efficient Qwen-VL backbones to balance capability and latency in 1-3B models.
Why learn
Compact models make VLA accessible: you can train and run them on a single consumer GPU or even on a Jetson. For anyone learning or prototyping with accessible platforms such as SO-100, these models are the most realistic entry point without giving up fundamental capabilities.
Key concepts
Knowledge distillation transfers capabilities from large models to small ones. SmolVLA uses async inference to separate perception and action, improving throughput. Aggressive quantization (INT4) and efficient backbones (Qwen2-VL) enable on-device inference at frequencies useful for control.
6 Comparison: size, license, embodiments, performance
What it is
A comparative analysis of the dimensions that matter when choosing a VLA: number of parameters, license type (closed vs. open vs. permissive), supported embodiments, control frequency, and success rate on benchmarks. There is no single best model; there is a best model for each set of constraints.
Why learn
Model selection is one of the most consequential decisions in a VLA project. Knowing how to compare trade-offs between RT-2 (powerful, closed), OpenVLA (balanced, open), π0 (high dexterity), and compact models (inexpensive, limited) avoids costly choices and rework.
Key concepts
Benchmarks such as LIBERO, SimplerEnv, and real-robot evaluations measure success rates on standardized tasks. Licenses range from proprietary (RT-2) to Apache 2.0 (OpenVLA, Octo). Embodiment coverage determines whether the model transfers to your specific robot without extensive fine-tuning.
Datasets: Open X-Embodiment and DROID
The data that powers VLAs — scale, diversity, and format.
1 Open X-Embodiment: the dataset that changed everything
What it is
Open X-Embodiment (OXE) is a collaborative effort bringing together more than 1 million robotic trajectories across 22 different embodiments, collected by 21 institutions around the world. It is the largest open robotics manipulation dataset and the training foundation for models like RT-X and OpenVLA.
Why learn
OXE is the dataset that made generalist VLAs possible. Without it, each lab would train isolated models on small, proprietary datasets. Understanding its structure and diversity is essential for knowing what a pretrained VLA already “knows” and where it is likely to fail.
Key concepts
Cross-embodiment means mixing data from robots with different morphologies (6- and 7-DoF arms, varied grippers) in a single training run. Differences in cameras, frequencies, and action spaces require normalization. This diverse scale is what enables generalization, but it also introduces noise and imbalance across sources.
2 DROID: in-the-wild diversity
What it is
DROID (Distributed Robot Interaction Dataset) includes about 76,000 trajectories collected across 564 scenes and 86 tasks, in highly diverse real-world environments (offices, kitchens, bathrooms). All use the same standardized Franka arm setup, making it easier to maintain consistency despite the diversity of scenes.
Why learn
DROID complements OXE by emphasizing environmental diversity, not robot diversity. This variety of scenes is exactly what VLAs need to generalize to new places. Understanding DROID helps you assess a model’s robustness outside the lab.
Key concepts
In-the-wild collection captures variations in lighting, backgrounds, and distractors that lab environments don’t have. Standardized hardware (Franka + fixed cameras) reduces confounding variables. Scene diversity improves spatial generalization, while the moderate number of tasks limits semantic generalization.
3 BridgeData V2: kitchen-scale manipulation
What it is
BridgeData V2 is a dataset with more than 60 thousand manipulation trajectories collected using a low-cost WidowX arm, focused on household kitchen tasks (moving objects, opening drawers, stacking). It is widely used as a benchmark for transfer and VLA fine-tuning.
Why learn
BridgeData shows that valuable data does not require expensive hardware. Because it uses an affordable arm and realistic tasks, it is frequently referenced in papers and is an excellent target for fine-tuning experiments when you do not have millions of your own trajectories.
Key concepts
Task diversity in a home environment provides a rich language-action signal. The low-cost WidowX arm demonstrates accessible and reproducible data collection. BridgeData is part of OXE, but is also used on its own as a benchmark for cross-model evaluation.
4 RLDS / TFDS format
What it is
RLDS (Reinforcement Learning Datasets) is the standardized format for storing robotic trajectories as sequences of episodes, each containing steps with an observation, action, reward, and metadata. It is built on TFDS (TensorFlow Datasets), enabling efficient streaming of enormous datasets.
Why learn
To train or fine-tune any modern VLA, you’ll need to read and write data in RLDS. It’s the format that brings OXE, DROID, and BridgeData together under a common interface, and mastering it is a practical prerequisite for any hands-on work with data.
Key concepts
The episode → steps structure naturally organizes sequential data for imitation learning. TFDS provides read pipelines with shuffling, prefetching, and parallelism. Normalizing action spaces and standardizing observation keys are critical steps for combining heterogeneous datasets.
5 Collection via teleoperation
What it is
The main way to generate robotic demonstration data: a human operator remotely controls the robot (via joystick, leader-follower arm, or VR) while the system records synchronized observations and actions. Each teleoperated demonstration becomes a labeled trajectory for imitation learning.
Why learn
Teleoperation is the cost and time bottleneck in building VLA datasets. Understanding its modes and limitations reveals why robotic data is so much scarcer than text data, and why data-efficiency techniques are so highly valued in the field.
Key concepts
Leader-follower (master-slave), as in ALOHA, replicates the movements of a manually operated leader arm on a follower arm. VR teleoperation maps the operator's hand to the end-effector. Demonstration quality (smoothness, lack of hesitation) directly affects the learned policy.
6 Quality vs. quantity and scaling laws
What it is
The debate over whether VLAs improve more with more data (quantity) or better-curated data (quality), and how performance scales with dataset size. For robotics, it is still unclear whether the same scaling laws observed in language and vision apply.
Why learn
This question determines where to invest resources: collect more trajectories, filter existing ones, or diversify sources. For a real project with a limited budget, knowing that quality often outweighs quantity can save months of teleoperation.
Key concepts
Data scaling laws relate dataset size to success rate, with diminishing returns. Data curation filters out low-quality demonstrations and rebalances overrepresented sources. Diversity (of tasks, objects, and scenes) often matters more than raw volume for generalization.
Hardware: Cobots, Humanoids, and Platforms
From collaborative arms to humanoids — and accessible platforms for learning.
1 Industrial cobots: Franka, UR, KUKA
What it is
Cobots (collaborative robots) are arms designed to operate safely alongside humans, without safety cages. Franka Emika (7 DoF, torque-sensitive), Universal Robots (UR3/UR5/UR10), and KUKA (LBR iiwa) are the most widely used platforms in research and light industry.
Why learn
Cobots are the most common hardware in VLA datasets and research (Franka dominates DROID). Knowing their capabilities, payload limits, and control interfaces is essential to understanding how demonstrations were collected and how to deploy a policy.
Key concepts
Degrees of freedom (DoF) define the arm's mobility; 7 DoF allow it to avoid singularities. Torque sensing at each joint enables force control and collision safety. Interfaces such as ROS, libfranka, and the real-time control protocol connect the arm to the AI model.
2 Humanoids: Figure, Unitree, Optimus, 1X
What it is
Bipedal humanoid robots designed to operate in environments built for humans. Figure 02, Unitree G1/H1, Tesla Optimus, and 1X's robots are the most advanced commercial platforms, combining bipedal locomotion, hand manipulation, and increasingly, control via VLAs.
Why learn
Humanoids are the most visible and ambitious frontier of VLA robotics. Understanding their challenges (balance, whole-body control, cost) helps separate hype from real capabilities and calibrate expectations about what VLAs can control today versus in a few years.
Key concepts
Whole-body control coordinates locomotion and manipulation simultaneously. Bipedal locomotion requires real-time dynamic balance control, typically delegated to controllers separate from the VLA. The Unitree G1 has drastically lowered the cost of entry, making humanoids accessible to research labs.
3 Grippers and end-effectors
What it is
The end-effector is the robot’s "hand," its final interface with the world. Options range from simple parallel grippers (two jaws that open and close) to anthropomorphic dexterous hands with multiple fingers and dozens of degrees of freedom, as well as suction cups and specialized tools.
Why learn
The end-effector determines which tasks are physically possible. A parallel gripper is enough for pick-and-place, but dexterous tasks (manipulating fabric, using tools) require complex hands. This choice directly affects the action space the VLA needs to control.
Key concepts
Parallel grippers have 1 DoF (open/close) and are robust and inexpensive. Dexterous hands (such as the Shadow Hand) have 20+ DoF, enabling fine manipulation but making control much harder. The end-effector's dimensionality determines the size of the action space at the VLA output.
4 Sensors: RGB-D, force/torque, tactile
What it is
The robot’s senses. RGB-D cameras (such as Intel RealSense) provide color and depth; force/torque sensors at the wrist measure interactions with objects; tactile sensors (GelSight, DIGIT) capture texture and contact at the fingertips. Together, they form the VLA’s perceptual input.
Why learn
Most current VLAs use only RGB, but fine-contact tasks (inserting a connector, handling a fragile object) depend on force and touch. Knowing what sensors are available reveals both the limitations of current models and directions for multimodal research.
Key concepts
RGB-D adds a depth channel that helps with 3D pose estimation. Force/torque sensing enables impedance control and contact detection. Tactile sensors provide high-resolution local feedback, which is crucial for dexterous manipulation but is still rarely integrated into mainstream VLAs.
5 Accessible platforms: SO-100, Koch, ALOHA
What it is
Low-cost, open-source robotic arms created to democratize robotics learning. SO-100/SO-101 (from Hugging Face's LeRobot project) cost a few hundred dollars; Koch and ALOHA use leader-follower setups for accessible teleoperation and data collection.
Why learn
These platforms are the realistic way for most people to start practicing VLA for real, with physical hardware rather than only in simulation. SO-100 + LeRobot is currently the recommended entry-level setup, integrated with ready-to-use models and datasets.
Key concepts
LeRobot is Hugging Face's library that brings together affordable hardware, datasets, and trainable policies. The leader-follower setup lets you teleoperate by moving a twin arm by hand. Its low cost makes it possible to collect your own data and fine-tune locally without an industrial investment.
6 Embedded compute: Jetson and latency
What it is
The compute hardware that runs the VLA, often embedded in the robot itself. The NVIDIA Jetson line (Orin Nano, AGX Orin) is the standard for on-device inference, offering a compact, low-power GPU suitable for autonomous deployment without an external server.
Why learn
The choice of compute determines whether the VLA runs in real time on the robot. A model that requires a datacenter GPU can’t go on an autonomous humanoid. Understanding Jetson’s latency and memory requirements is critical to designing systems that can actually be deployed.
Key concepts
Inference latency must fit within the control period (e.g., <100 ms for 10 Hz). Quantization (INT8/INT4) and TensorRT optimize the model for Jetson. The edge vs. cloud trade-off pits autonomy and low on-device latency against the greater capacity of remote servers, with the risk of network delays.
Safety and Standards: ISO 10218:2025
Safety-by-design and the standards governing human-robot collaboration.
1 ISO 10218:2025: the industrial robot standard
What it is
ISO 10218:2025 is the international standard that defines safety requirements for industrial robots and their integration. The 2025 revision unified and updated parts 1 (robot) and 2 (system/cell integration), explicitly incorporating collaborative operation and modern human-robot interaction scenarios.
Why learn
Any real-world robot deployment in a production environment must comply with this standard. For those working with VLAs, it defines the safety envelope within which the model can operate, making it a legal and ethical prerequisite before putting a robot near people.
Key concepts
The standard specifies design requirements, protective stops, and safety functions with a performance level (PL) defined by ISO 13849. It distinguishes between a robot (Part 1) and an integrated cell (Part 2). The 2025 edition aligns with the evolution of cobots and continuous collaborative operation.
2 ISO/TS 15066: force and pressure limits
What it is
ISO/TS 15066 is the technical specification that supplements ISO 10218 with a specific focus on collaborative robots. It defines biomechanical force and pressure limits for different regions of the human body, ensuring that robot-human contact does not cause injury.
Why learn
It’s the standard that makes human-robot collaboration physically safe. Any VLA controlling a cobot operating near people must respect these limits, which in practice restrict the maximum speed and force of the actions the model can issue.
Key concepts
Biomechanical limits are tabulated by body part (e.g., maximum pressure on the hand vs. the face). Defines the four modes of collaboration, including power-and-force limiting (PFL) and speed-and-separation monitoring (SSM). It is a technical specification (TS), intended as guidance, but referenced by ISO 10218.
3 Safety-by-design for VLA
What it is
The philosophy of embedding safety from the system’s design stage, instead of adding it as a patch later. For VLAs, this means designing the architecture, action space, and model limits so that dangerous behaviors are structurally difficult or impossible.
Why learn
VLAs are unpredictable generative models; you cannot simply trust that they "will behave well." Safety by design is what makes it possible to use probabilistic models in physical systems without exposing people to risk, making it a core competency for any responsible robotics engineer.
Key concepts
Action clamping physically limits the model's output to safe speed and force ranges. Model-independent safety layers (deterministic safety controllers) act as a safety net. The principle of defense-in-depth combines multiple redundant barriers so that no single failure causes harm.
4 Emergency stop and zones: SSM and PFL
What it is
The physical and logical mechanisms that stop or modulate a robot’s movement in the presence of humans. Includes emergency stop buttons (e-stop), monitored safety zones, and the two main collaborative modes: SSM (speed and separation monitoring) and PFL (power and force limiting).
Why learn
These are the last-resort safeguards that protect people regardless of what the VLA decides to do. Understanding how SSM and PFL work reveals where the AI model ends and the deterministic safety system takes over.
Key concepts
Speed-and-separation monitoring (SSM) reduces speed as a human approaches, stopping before contact. Power-and-force limiting (PFL) allows contact, but within the biomechanical limits of TS 15066. Category 0/1 emergency stop safely cuts power, independently of the control software.
5 Risk analysis and mitigation
What it is
The systematic process of identifying potential hazards in a robotic system, estimating the severity and likelihood of each, and defining measures to reduce them to acceptable levels. It is a mandatory step required by standards before any deployment.
Why learn
Every compliant robotics integration starts with a risk assessment. For systems with VLA, it needs to account for the model’s unpredictability, broadening the scope of hazards to evaluate. Knowing how to conduct one is a prerequisite for any serious project.
Key concepts
The mitigation hierarchy prioritizes eliminating hazards, followed by engineering protections, then warnings and PPE. Methods such as HAZOP and FMEA structure failure identification. Risk reduction is iterative: assess, mitigate, and reassess until the residual risk is acceptable.
6 Guardrails, fail-safe, and VLA validation
What it is
The set of safeguards specific to generative models in physical control: guardrails that filter or block unsafe actions proposed by the VLA, fail-safe behaviors that bring the robot to a safe state in the event of a failure, and verification and validation (V&V) processes that demonstrate the system’s behavior.
Why learn
It’s the boundary between experimental and deployable VLAs. Since generative models can hallucinate actions, guardrails and fail-safes are what make them trustworthy in production. Without robust V&V, no VLA system should operate near people.
Key concepts
Guardrails validate each action against safety constraints before execution, rejecting commands outside the envelope. Fail-safe behaviors define the default safe state (stop, gently release the load) when something goes wrong. Verification and validation combine simulation tests, runtime monitoring, and statistical evaluation of behavior at scale.