MAROKO133 Breaking ai: ByteDance Introduces Astra: A Dual-Model Architecture for Autonomou

📌 MAROKO133 Breaking ai: ByteDance Introduces Astra: A Dual-Model Architecture for

The increasing integration of robots across various sectors, from industrial manufacturing to daily life, highlights a growing need for advanced navigation systems. However, contemporary robot navigation systems face significant challenges in diverse and complex indoor environments, exposing the limitations of traditional approaches. Addressing the fundamental questions of “Where am I?”, “Where am I going?”, and “How do I get there?”, ByteDance has developed Astra, an innovative dual-model architecture designed to overcome these traditional navigation bottlenecks and enable general-purpose mobile robots.

Traditional navigation systems typically consist of multiple, smaller, and often rule-based modules to handle the core challenges of target localization, self-localization, and path planning. Target localization involves understanding natural language or image cues to pinpoint a destination on a map. Self-localization requires a robot to determine its precise position within a map, especially challenging in repetitive environments like warehouses where traditional methods often rely on artificial landmarks (e.g., QR codes). Path planning further divides into global planning for rough route generation and local planning for real-time obstacle avoidance and reaching intermediate waypoints.

While foundation models have shown promise in integrating smaller models to tackle broader tasks, the optimal number of models and their effective integration for comprehensive navigation remained an open question.

ByteDance’s Astra, detailed in their paper “Astra: Toward General-Purpose Mobile Robots via Hierarchical Multimodal Learning” (website: https://astra-mobility.github.io/), addresses these limitations. Following the System 1/System 2 paradigm, Astra features two primary sub-models: Astra-Global and Astra-Local. Astra-Global handles low-frequency tasks like target and self-localization, while Astra-Local manages high-frequency tasks such as local path planning and odometry estimation. This architecture promises to revolutionize how robots navigate complex indoor spaces.

Astra-Global: The Intelligent Brain for Global Localization

Astra-Global serves as the intelligent core of the Astra architecture, responsible for critical low-frequency tasks: self-localization and target localization. It functions as a Multimodal Large Language Model (MLLM), adept at processing both visual and linguistic inputs to achieve precise global positioning within a map. Its strength lies in utilizing a hybrid topological-semantic graph as contextual input, allowing the model to accurately locate positions based on query images or text prompts.

The construction of this robust localization system begins with offline mapping. The research team developed an offline method to build a hybrid topological-semantic graph G=(V,E,L):

• V (Nodes): Keyframes, obtained by temporal downsampling of input video and SfM-estimated 6-Degrees-of-Freedom (DoF) camera poses, act as nodes encoding camera poses and landmark references.
• E (Edges): Undirected edges establish connectivity based on relative node poses, crucial for global path planning.
• L (Landmarks): Semantic landmark information is extracted by Astra-Global from visual data at each node, enriching the map’s semantic understanding. These landmarks store semantic attributes and are connected to multiple nodes via co-visibility relationships.

In practical localization, Astra-Global’s self-localization and target localization capabilities leverage a coarse-to-fine two-stage process for visual-language localization. The coarse stage analyzes input images and localization prompts, detects landmarks, establishes correspondence with a pre-built landmark map, and filters candidates based on visual consistency. The fine stage then uses the query image and coarse output to sample reference map nodes from the offline map, comparing their visual and positional information to directly output the predicted pose.

For language-based target localization, the model interprets natural language instructions, identifies relevant landmarks using their functional descriptions within the map, and then leverages landmark-to-node association mechanisms to locate relevant nodes, retrieving target images and 6-DoF poses.

To empower Astra-Global with robust localization abilities, the team employed a meticulous training methodology. Using Qwen2.5-VL as the backbone, they combined Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO). SFT involved diverse datasets for various tasks, including coarse and fine localization, co-visibility detection, and motion trend estimation. In the GRPO phase, a rule-based reward function (including format, landmark extraction, map matching, and extra landmark rewards) was used to train for visual-language localization. Experiments showed GRPO significantly improved Astra-Global’s zero-shot generalization, achieving 99.9% localization accuracy in unseen home environments, surpassing SFT-only methods.

Astra-Local: The Intelligent Assistant for Local Planning

Astra-Local acts as the intelligent assistant for Astra’s high-frequency tasks, a multi-task network capable of efficiently generating local paths and accurately estimating odometry from sensor data. Its architecture comprises three core components: a 4D spatio-temporal encoder, a planning head, and an odometry head.

The 4D spatio-temporal encoder replaces traditional mobile stack perception and prediction modules. It begins with a 3D spatial encoder that processes N omnidirectional images through a Vision Transformer (ViT) and Lift-Splat-Shoot to convert 2D image features into 3D voxel features. This 3D encoder is trained using self-supervised learning via 3D volumetric differentiable neural rendering. The 4D spatio-temporal encoder then builds upon the 3D encoder, taking past voxel features and future timestamps as input to predict future voxel features through ResNet and DiT modules, providing current and future environmental representations for planning and odometry.

The planning head, based on pre-trained 4D features, robot speed, and task information, generates executable trajectories using Transformer-based flow matching. To prevent collisions, the planning head incorporates a masked ESDF loss (Euclidean Signed Distance Field). This loss calculates the ESDF of a 3D occupancy map and applies a 2D ground truth trajectory mask, significantly reducing collision rates. Experiments demonstrate its superior performance in collision rate and overall score on out-of-distribution (OOD) datasets compared to other methods.

The odometry head predicts the robot’s relative pose using current and past 4D features and additional sensor data (e.g., IMU, wheel data). It trains a Transformer model to fuse information from different sensors. Each sensor modality is processed by a specific tokenizer, combined with modality embeddings and temporal positional embeddi…

Konten dipersingkat otomatis.

🔗 Sumber: syncedreview.com

📌 MAROKO133 Breaking ai: AWS launches Kiro powers with Stripe, Figma, and Datadog

Amazon Web Services on Wednesday introduced Kiro powers, a system that allows software developers to give their AI coding assistants instant, specialized expertise in specific tools and workflows — addressing what the company calls a fundamental bottleneck in how artificial intelligence agents operate today.

AWS made the announcement at its annual re:Invent conference in Las Vegas. The capability marks a departure from how most AI coding tools work today. Typically, these tools load every possible capability into memory upfront — a process that burns through computational resources and can overwhelm the AI with irrelevant information. Kiro powers takes the opposite approach, activating specialized knowledge only at the moment a developer actually needs it.

"Our goal is to give the agent specialized context so it can reach the right outcome faster — and in a way that also reduces cost," said Deepak Singh, Vice President of Developer Agents and Experiences at Amazon, in an exclusive interview with VentureBeat.

The launch includes partnerships with nine technology companies: Datadog, Dynatrace, Figma, Neon, Netlify, Postman, Stripe, Supabase, and AWS's own services. Developers can also create and share their own powers with the community.

Why AI coding assistants choke when developers connect too many tools

To understand why Kiro powers matters, it helps to understand a growing tension in the AI development tool market.

Modern AI coding assistants rely on something called the Model Context Protocol, or MCP, to connect with external tools and services. When a developer wants their AI assistant to work with Stripe for payments, Figma for design, and Supabase for databases, they connect MCP servers for each service.

The problem: each connection loads dozens of tool definitions into the AI's working memory before it writes a single line of code. According to AWS documentation, connecting just five MCP servers can consume more than 50,000 tokens — roughly 40 percent of an AI model's context window — before the developer even types their first request.

Developers have grown increasingly vocal about this issue. Many complain that they don't want to burn through their token allocations just to have an AI agent figure out which tools are relevant to a specific task. They want to get to their workflow instantly — not watch an overloaded agent struggle to sort through irrelevant context.

This phenomenon, which some in the industry call "context rot," leads to slower responses, lower-quality outputs, and significantly higher costs — since AI services typically charge by the token.

Inside the technology that loads AI expertise on demand

Kiro powers addresses this by packaging three components into a single, dynamically-loaded bundle.

The first component is a steering file called POWER.md, which functions as an onboarding manual for the AI agent. It tells the agent what tools are available and, crucially, when to use them. The second component is the MCP server configuration itself — the actual connection to external services. The third includes optional hooks and automation that trigger specific actions.

When a developer mentions "payment" or "checkout" in their conversation with Kiro, the system automatically activates the Stripe power, loading its tools and best practices into context. When the developer shifts to database work, Supabase activates while Stripe deactivates. The baseline context usage when no powers are active approaches zero.

"You click a button and it automatically loads," Singh said. "Once a power has been created, developers just select 'open in Kiro' and it launches the IDE with everything ready to go."

How AWS is bringing elite developer techniques to the masses

Singh framed Kiro powers as a democratization of advanced development practices. Before this capability, only the most sophisticated developers knew how to properly configure their AI agents with specialized context — writing custom steering files, crafting precise prompts, and manually managing which tools were active at any given time.

"We've found that our developers were adding in capabilities to make their agents more specialized," Singh said. "They wanted to give the agent some special powers to do a specific problem. For example, they wanted their front end developer, and they wanted the agent to become an expert at backend as a service."

This observation led to a key insight: if Supabase or Stripe could build the optimal context configuration once, every developer using those services could benefit.

"Kiro powers formalizes that — things that people, only the most advanced people were doing — and allows anyone to get those kind of skills," Singh said.

Why dynamic loading beats fine-tuning for most AI coding use cases

The announcement also positions Kiro powers as a more economical alternative to fine-tuning, the process of training an AI model on specialized data to improve its performance in specific domains.

"It's much cheaper," Singh said, when asked how powers compare to fine-tuning. "Fine-tuning is very expensive, and you can't fine-tune most frontier models."

This is a significant point. The most capable AI models from Anthropic, OpenAI, and Google are typically "closed source," meaning developers cannot modify their underlying training. They can only influence the models' behavior through the prompts and context they provide.

"Most people are already using powerful models like Sonnet 4.5 or Opus 4.5," Singh said. "What those models need is to be pointed in the right direction."

The dynamic loading mechanism also reduces ongoing costs. Because powers only activate when relevant, developers aren't paying for token usage on tools they're not currently using.

Where Kiro powers fits in Amazon's bigger bet on autonomous AI agents

Kiro powers arrives as part of a broader push by AWS into what the company calls "agentic AI" — artificial intelligence systems that can operate autonomously over extended periods.

Earlier at re:Invent, AWS announced three "frontier agents" designed to work for hours or days without human intervention: the Kiro autonomous agent for software development, the AWS security agent, and the AWS DevOps agent. These represent a different approach from Kiro powers — tackling large, ambiguous problems rather than providing specialized expertise for specific tasks.

The two approaches are complementary. Frontier agents handle complex, multi-day projects that require autonomous decision-making across multiple codebases. Kiro powers, by contrast, gives developers precise, efficient tools for everyday development tasks where speed and token efficiency matter most.

The company is betting that developers need both ends of this spectrum to be productive.

…

Konten dipersingkat otomatis.

🔗 Sumber: venturebeat.com

---