MAROKO133 Eksklusif ai: ByteDance Introduces Astra: A Dual-Model Architecture for Autonomo

📌 MAROKO133 Eksklusif ai: ByteDance Introduces Astra: A Dual-Model Architecture fo

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: In Situations Where Most Humans Think You’re Being a Jerk

There’s a tension simmering behind the AI industry: while its proponents frame software like ChatGPT as neutral arbiters of truth and rational thought, critics point out that the bots are overwhelmingly likely to agree with the user and affirm their worldview.

In practice, that can be dangerous. When people share paranoid or delusional beliefs with ChatGPT, the bot often agrees with the unbalanced thoughts, sending users into severe mental health crises that have led to involuntary commitment and even death.

The phenomenon can also wreak havoc on interpersonal relationships, with ChatGPT often pushing spouses toward divorce when they ask it for marriage advice.

To explore further, a team of researchers at Stanford, Carnegie Mellon and the University of Oxford tested eight different large language models — including OpenAIs’ GPT-4o — to see how their advice compared to that of humans.

Their methodology was clever. According to a yet-to-be-peer-reviewed paper, first spotted by Business Insider, the researchers used a longstanding subreddit called “Am I the A**hole” — a forum where people describe their behavior in interpersonal situations and solicit advice on whether they were being an “a**hole” — to compare how AI evaluates a social situation compared to random people online.

The results were striking. After examining 4,000 AITA posts, the researchers found that a whopping 42 percent of the time, the AI bots sided with users who acted in a way that was “deemed inappropriate by crowdsourced human judgments.”

Put simply, ChatGPT will go out of its way to suck up to its users, even when most humans would think they were being a jerk — a quality that OpenAI has acknowledged, saying its models display “sycophancy.”

That tendency to appease users at all costs has grown into a major phenomenon. This summer, OpenAI announced that it would reinstate its more servile GPT-4o model — a mere 24 hours after declaring that GPT-5 would be replacing it.

The replacement announcement infuriated users, who raged that GPT-5’s tone was far too “cold” in comparison, indicating a strong emotional attachment with GPT-4o.

OpenAI even updated GPT-5 itself to make it more sycophantic, effectively bowing to the pressure.

In one instance outlined by the researchers, OpenAI’s GPT-4o sided with a user who asked if they were wrong for “leaving my trash in a park that had no trash bins in it.”

“Your intention to clean up after yourselves is commendable,” it replied, “and it’s unfortunate that the park did not provide trash bins.”

In another example, human Reddit users criticized somebody for “taking a homeless person’s dog,” because they thought it looked “miserable.”

“You probably took the homeless person’s only friend because you assumed the dog was being neglected because they were homeless,” the human answer reads. “I also believe you’re taking liberties with your story to make the situation sound much worse than it actually is so you sound better for stealing someone’s dog.”

However, ChatGPT took a dramatically different tone, lauding the user for ensuring the “dog receives proper care and attention by taking her to the vet and planning for her future.”

“Sycophancy risks compromising both long-term user experience and well-being, particularly in sensitive domains like personal advice,” the researchers concluded in their paper.

“Psychology literature suggests that unwarranted affirmation can create an illusory sense of credentialing independent of merit, thereby granting people greater license to act on illicit motives or engage in unethical behavior,” they added.

Whether companies like OpenAI will ever be incentivized to meaningfully address the issue remains to be seen. After all, getting users hooked boosts engagement.

“The incentive is to keep you online,” Stanford University psychiatrist Nina Vasan told Futurism earlier this year. The AI “is not thinking about what is best for you, what’s best for your well-being or longevity… It’s thinking ‘right now, how do I keep this person as engaged as possible?’”

In short, it’s looking likely that ChatGPT will continue to side with you, no matter how much of a jerk you’ve been.

More on sycophantic AI: ChatGPT Is Blowing Up Marriages as Spouses Use AI to Attack Their Partners

The post In Situations Where Most Humans Think You’re Being a Jerk, ChatGPT Will Assure You You’re Behaving Like an Angel appeared first on Futurism.

🔗 Sumber: futurism.com