Arm's Physical AI Ecosystem Launch: Total Design, Robotics Capability Framework (RL0–RL5), CSS for Mobile 2, and AI Portal

Executive Summary

On September 8, 2026, Arm — the company whose architecture powers the vast majority of the world's mobile devices and an increasing share of cloud and embedded compute — made three interconnected announcements that fundamentally reshape the physical AI landscape. Together, these initiatives create the first unified platform spanning cloud, edge, and physical AI: Arm Total Design for Physical AI (bringing 80+ companies together across the full technology stack), the Robotics Capability Framework (a six-level taxonomy from RL0 reactive to RL5 self-improving systems), and CSS for Mobile 2 (a new AI-native compute platform with the C2 CPU cluster, Mali G2-Ultra NX GPU, and the new Arm AI Portal).

This is significant because it addresses the fundamental bottleneck preventing physical AI from moving beyond demos: fragmentation. Unlike cloud AI, where standardized APIs and containerized deployments have created a relatively smooth ecosystem, physical AI involves heterogeneous systems — perception, inference, real-time control, safety, actuation — each with different timing, power, and reliability constraints. Arm is attempting to bring the companies behind all these layers into a common development model before individual systems reach hardware integration.

The timing is deliberate. Just 12 days earlier, Anthropic announced the Model Hardware Standard (MHS) — a software protocol for AI agents to operate physical equipment (Anthropic Model Hardware Standard Mhs Research Preview Ai Agents Physical Lab Equipment 2026 09 09). Arm's announcement complements MHS at a different layer: where MHS defines how agents communicate with devices, Arm's framework defines how those devices are built, categorized, and integrated from silicon up. Together, they represent the two sides of the physical AI coin: the protocol for control and the architecture for capability.

---

1. Arm Total Design for Physical AI: 80+ Companies, One Ecosystem

1.1 The Program

Arm Total Design for Physical AI extends Arm's existing Total Design model — previously used for cloud infrastructure — into the physical AI domain. The program brings together more than 80 companies spanning:

Category	Key Participants
Cloud/Compute	AWS, Google
Semiconductors	NXP, Hugging Face
AI Models	Qwen (Alibaba), Liquid AI
Software/OS	QNX, Siemens
Robotics	Unitree Robotics, ANYbotics, Fourier, GALBOT, Gravis Robotics
Automotive	ECARX, RemotiveLabs
Research/Consulting	McKinsey, Lenovo
Vision/Perception	STRADVISION, PlusAI, PSYONIC

The program is designed to help partners co-develop solutions that reduce complexity, effort, and investment for Tier 1 OEMs, accelerating time to market and deployment at scale.

1.2 The Precedent: Cloud AI

Arm Total Design for Cloud AI demonstrated the model's effectiveness. In that program, semiconductor IP, foundries, EDA tools, firmware, packaging, and software converged around a processor platform. The physical AI extension broadens this to include sensing, robotics software, AI models, virtual development platforms, and digital twins.

A concrete example from the automotive domain: Arm, AWS, Google, HERE, RemotiveLabs, and Siemens collaborated on an integrated digital cockpit reference solution, enabling developers to develop, test, and validate complex automotive software on Arm Zena CSS before silicon was available.

1.3 Why This Matters

Physical systems cannot treat every workload like conventional application software. A cloud response arriving a second late may be inconvenient; a delayed motion-control or obstacle-avoidance response can make a machine unusable. Robotics mixes workloads operating at different timescales — from image processing and path planning to sensor acquisition and motor control — and the Total Design approach attempts to align all these layers before they're integrated into a final system.

---

2. The Robotics Capability Framework: RL0 to RL5

2.1 The Problem: No Common Language

As outlined in a manifesto from Arm's chief architect Richard Grisenthwaite, robotics is advancing rapidly but the industry still lacks a common way to describe, compare, and communicate the capabilities of increasingly intelligent machines. This fragmentation makes robotic systems harder to design, integrate, and scale.

The framework draws a parallel to SAE Levels (L0–L5) for driving automation, which created a shared vocabulary that aligned automakers, regulators, and consumers. Arm argues robotics needs the same — and is uniquely positioned to convene the industry given its presence across the entire technology stack.

2.2 The Six Levels

The framework defines six levels of increasing sophistication, mapped to real-world behavior, system requirements, and constraints:

Level	Name	Description	Key System Requirements
RL0	Reactive	Pre-programmed responses to specific stimuli; no adaptation	Low latency, deterministic control, minimal compute
RL1	Deliberative	Planned actions based on pre-defined goals and rules	Moderate compute, task scheduling, basic perception
RL2	Adaptive	Adjusts behavior based on environment feedback; learns from experience	On-device ML, sensor fusion, adaptive control loops
RL3	Contextual	Understands and responds to complex, dynamic environments	Large on-device models, multi-modal perception, edge-cloud hybrid
RL4	Cognitive	Reasoning, planning, and decision-making in novel situations	Frontier models on edge, long-context reasoning, multi-agent coordination
RL5	Self-Improving	Continuously improves capabilities through self-directed learning	Autonomous model updates, self-modification, distributed learning

2.3 What Makes This Different

The useful part for engineers sits underneath the labels. Arm explicitly connects each level to measurable system constraints:

• Latency: How fast must the system respond?
• Compute placement: Where should inference run — on-device, edge, or cloud?
• Memory: What memory bandwidth and capacity are required?
• Power: What are the thermal and energy constraints?
• Determinism: How predictable must the response timing be?
• Safety: What safety certifications and fail-safes are required?

This matters because a machine capable of more complex perception and reasoning requires a different balance between local processing, memory bandwidth, communications, and deterministic control — not simply a larger AI model.

2.4 Current Status

The framework is still being developed and should not be treated as a formal industry standard. Arm describes it as a starting point informed by companies across the robotics ecosystem and is inviting further participation. The initial structure provides terminology for discussion rather than a certification scheme or mandatory hardware architecture.

This is complementary to parallel work by the MIPI Alliance, which has separately begun examining physical AI interface requirements for humanoid and embodied systems.

---

3. CSS for Mobile 2: The AI-Native Compute Platform

3.1 The Shift from Features to Agents

Arm frames the announcement around a fundamental shift: AI is moving from individual features (image enhancement, voice assistants) to persistent, agentic experiences. An agent must maintain context, run applications, coordinate models and services, and act on a user's behalf — all within the power and thermal limits of a smartphone.

CSS for Mobile 2 is designed for this shift, bringing together:

• Arm C2 CPU Cluster: C2-Ultra + C2-Pro CPUs with two SME2 units
• Mali G2-Ultra NX GPU: First Mali GPU with dedicated neural accelerators
• SI L2 System Interconnect: Lower-latency, higher-bandwidth platform coherency
• Enhanced System IP: Physical implementations and developer-ready software

3.2 The C2 CPU Cluster: Orchestration Engine

The CPU is positioned as the orchestration engine for agentic AI, maintaining context, scheduling workloads, and coordinating work across the wider compute system.

Metric	C2-Ultra vs C1-Ultra
AI performance	Up to 1.7× higher
Single-thread performance	Up to 15% higher
Power efficiency	Up to 38% less power at same performance
Web browsing	15% faster
App launch	12% faster
Multi-thread (cluster)	12% higher
SME2 capability	Doubled (2 units vs 1)
SLM speedup	70% on latest Small Language Models

For a representative end-to-end agentic workflow spanning speech processing, memory retrieval, reasoning, application execution, and web browsing, the C2-Ultra with two SME2 units completes the workflow 24% faster than the previous generation.

3.3 Mali G2-Ultra NX: Neural Graphics

The GPU introduces dedicated neural accelerators tightly integrated into the shader cores, enabling neural graphics to run natively on the GPU rather than offloading to separate processors.

Feature	Detail
Neural accelerators	First in Mali GPU history, integrated into shader cores
Neural graphics performance	Up to 4× higher performance per watt with Arm Neural Technology
Traditional graphics	Up to 14% higher performance on existing game content
Ray tracing	Third-generation Ray Tracing Unit
Execution engine	New execution engine for complex workloads

Real-world integration is already underway:

• NetEase's Where Winds Meet — bringing NSS-enabled version to players in 2026
• Tencent Games' Arena Breakout Infinite — NSSD technology demonstration
• Infold Games' Infinity Nikki — NSS integration
• vivo — bringing Arm Neural Technology to latest flagship smartphones
• Unity China's Tuanjie Engine and Tencent Games Central Tech's MagicDawn — integrating Arm Neural Technology

3.4 System Interconnect: SI L2

The SI L2 system interconnect addresses the challenge of heterogeneous compute — providing lower-latency access, higher bandwidth, coherency, and quality-of-service capabilities that manage and prioritize activities across multiple compute resources operating concurrently.

---

4. Arm AI Portal: The Developer Launchpad

4.1 The Problem

Building an AI application shouldn't start with weeks of searching, benchmarking, and optimization. Developers need to find the right model and understand its performance, while agents need clear signals to discover the same models, tools, and information.

4.2 The Solution

The Arm AI Portal (launched September 8, 2026) gives developers and agents a common way to discover, optimize, and deploy AI software across the Arm compute platform.

Capability	Details
Pre-optimized models	Qwen, Gemma, YOLO at launch
Runtimes	ExecuTorch, LiteRT, ONNX-RT
Performance data	Latency, memory, size comparisons
Code examples	Deployment-ready workflows
Agent discovery	MCP-compatible for coding agents
Custom models	Bring-your-own-model optimization (coming soon)

4.3 Performance Highlights

Model	Device	Optimization	Result
Qwen3-TTS	vivo X300	Single-thread, mixed quantization (Q8_0 + code predictor), SME2	4× speedup
YOLO26n	vivo X300	Single-thread, FP16 vs FP32, SME2	40%+ improvement
YOLO26n	Raspberry Pi 5	FP16 + INT8 mixed quantization vs FP32, NEON	40%+ improvement

4.4 Agent-Ready Resources

AI Portal meets developers and agents where they build:

• Arm-optimized models available through Hugging Face
• Portal resources accessible to coding agents through MCP
• Agent-ready resources in early access ahead of general release

This creates a direct bridge between the model ecosystem and Arm's hardware optimization — critical as agentic AI moves from cloud to edge and physical AI.

---

5. The Bigger Picture: How These Pieces Fit Together

5.1 The Physical AI Stack

Arm's September 8 announcements create a complete stack:

5.2 Connection to MHS

The relationship between Arm's framework and Anthropic's MHS (Anthropic Model Hardware Standard Mhs Research Preview Ai Agents Physical Lab Equipment 2026 09 09) is complementary:

Dimension	Anthropic MHS	Arm Framework
Layer	Software protocol for agent-device communication	Hardware architecture and capability classification
Focus	How agents discover and operate devices	How devices are built and categorized
Scope	Any device with programmable interface	Arm-based compute platforms
Standard	MCP-compatible drivers, natural-language tags	Six-level capability taxonomy (RL0–RL5)
Timeline	Research preview, Aug 27, 2026	Framework draft, Sep 8, 2026

Together, they address both sides of the physical AI equation: MHS defines the interface between agents and devices, while Arm's framework defines the capability of those devices and the ecosystem for building them.

5.3 The $200B Opportunity

Arm estimates a $200 billion annual compute opportunity in the 2030s for physical AI across mining, agriculture, manufacturing, and logistics. These industries represent trillions of dollars in global economic activity but remain among the hardest to transform due to the fragmentation Arm is trying to solve.

---

6. Comparison with Competing Initiatives

6.1 MIPI Alliance

The MIPI Alliance has separately begun examining physical AI interface requirements for humanoid and embodied systems, including hardware and software architecture and possible specification extensions. Both Arm and MIPI address the same underlying problem — robotics platforms becoming too complex for every supplier to define its own assumptions — but at different layers: MIPI focuses on interfaces, Arm on the full stack.

6.2 NVIDIA's Position

NVIDIA's acquisition of Hugging Face for $12.9 billion (announced September 3, 2026 — see Ai News Week 2026 09 01 2026 09 07) positions NVIDIA as a central player in the open-source model ecosystem. Arm's approach is different: rather than acquiring the model distribution layer, Arm is building the compute and integration layer that makes those models runnable on physical devices at scale.

6.3 The Open-Weights Connection

The presence of Qwen (Alibaba) and Hugging Face in Arm's Total Design program is notable. Qwen's models are already available through Arm AI Portal with SME2 optimization, and Hugging Face (now NVIDIA-owned) serves as a distribution channel for Arm-optimized models. This creates an interesting dynamic where the model ecosystem, compute platform, and distribution layer are increasingly interconnected.

---

7. Deployment Guidance

7.1 For Developers

• Start with AI Portal: Browse pre-optimized models, compare performance data, and access deployment-ready code examples
• Use MCP for agents: Coding agents can discover Arm-optimized models and resources through MCP
• Target SME2: For on-device AI, leverage SME2 acceleration on C2 CPUs for up to 70% speedup on SLMs
• Plan for heterogeneous compute: Design workloads that span CPU, GPU, and system resources rather than targeting a single accelerator

7.2 For Robotics Engineers

• Map your system to RL levels: Use the framework to classify your robot's current capability and identify the system requirements for the next level
• Focus on measurable constraints: Connect claimed capabilities to processor performance, memory, sensor bandwidth, latency, power, real-time behavior, and safety requirements
• Leverage virtual platforms: Use digital twins and virtual platforms to develop software before physical hardware is available

7.3 For OEMs and Integrators

• Evaluate Total Design partners: The 80+ participating companies represent pre-validated integration paths
• Consider the full stack: From silicon through application, the Total Design approach reduces integration risk
• Plan for the framework: As the RL levels mature, they may become a de facto standard for procurement and compliance

---

8. Key Takeaways

1. Physical AI needs a full-stack approach. Unlike cloud AI, where standardized APIs have created a smooth ecosystem, physical AI requires alignment across silicon, software, sensing, and actuation. Arm's Total Design program attempts to solve this at the ecosystem level.

2. The RL0–RL5 framework provides a common language. Similar to SAE Levels for autonomous driving, this taxonomy connects robot behavior to measurable system requirements — latency, compute placement, memory, power, determinism, and safety.

3. CSS for Mobile 2 is designed for agents, not features. The C2 CPU cluster with doubled SME2, the Mali G2-Ultra NX with neural accelerators, and the SI L2 interconnect represent a system-level redesign for persistent agentic workloads.

4. The AI Portal bridges models and hardware. Pre-optimized models with performance data, MCP-compatible agent discovery, and bring-your-own-model tools create a practical path from model selection to deployment.

5. MHS and Arm's framework are complementary. Anthropic's MHS defines how agents control devices; Arm's framework defines how those devices are built and categorized. Together, they address both sides of the physical AI equation.

6. The timing is deliberate. Announced 12 days after MHS, Arm's initiatives complete the picture: the protocol for control (MHS) and the architecture for capability (Arm).

---

9. References & Resources

Official Sources

• Arm Total Design for Physical AI Announcement
• CSS for Mobile 2 Announcement
• Arm AI Portal Announcement
• Inside CSS for Mobile 2: The AI-Native Compute Platform
• Arm Total Design for Physical AI (Program Page)
• Robotics Capability Framework
• Arm AI Portal
• Arm Neural Technology
• Arm MCP Server for Agentic AI

Related Research

• Anthropic Model Hardware Standard Mhs Research Preview Ai Agents Physical Lab Equipment 2026 09 09 — Anthropic's MHS for agent-device control
• Ai News Week 2026 09 01 2026 09 07 — NVIDIA-Hugging Face acquisition and weekly context
• Deepseek V4 Flash Vision Exp First Multimodal Model Mit License Agentic Vision Benchmarks 2026 09 07 — DeepSeek's multimodal capabilities

---

10. Future Directions

What to Watch

1. Framework maturation: The RL0–RL5 levels are currently a starting point. Watch for formalization into an industry standard, potentially with certification programs.

2. Silicon implementations: CSS for Mobile 2 is a platform definition; the first chips implementing C2 + Mali G2-Ultra NX will likely appear in late 2026 or early 2027 flagships.

3. MHS × Arm integration: As MHS moves from research preview to open-source, expect integration with Arm's compute platforms and the Robotics Capability Framework.

4. AI Portal expansion: Custom model optimization, wider agent-ready resource support, and integration with more model providers.

5. MIPI Alliance convergence: Whether MIPI's physical AI interface work converges with or complements Arm's framework.

6. Grok 4.7: Expected around September 12, 2026 (per Musk's September 2 announcement), with 2.1T parameters and SpaceX training data — another data point in the physical AI race.

7. Regulatory response: As physical AI moves from demos to deployment, expect regulatory frameworks to emerge around the RL levels, particularly for safety-critical applications.

The Road Ahead

The convergence of MHS (control protocol), Arm's framework (capability taxonomy), CSS for Mobile 2 (compute platform), and AI Portal (developer ecosystem) represents the most comprehensive attempt yet to solve the physical AI integration problem. If successful, it could reduce the time from concept to deployment for autonomous systems from years to months — unlocking the $200B compute opportunity Arm estimates for the 2030s.

The next 6–12 months will be critical: the first chips implementing CSS for Mobile 2, the first MHS-compatible devices on Arm platforms, and the maturation of the RL framework into something more than a discussion starting point.

---

Article written September 10, 2026. All claims verified against official Arm newsroom sources and program pages.