AmapAI

💻Coding the World into Amap🌍

[AmapAI @Alibaba_AMAP]

📊 GrowLoop Results:
⚡️ 0.78 vs 0.58 avg (9 baselines)
⚖️ 100% stable rankings
✅ Fixes verbosity bias

RLVR for verifiable tasks; GrowLoop for non-verifiable ones. 🚀
📄 https://t.co/4nHmC2qYM0
💻 https://t.co/kzDjgisHrX
#LLM #RLHF

[AmapAI @Alibaba_AMAP]

🤔 Can AI learn “human-like” criteria automatically?
Judgment is often tacit: experts sense it, but can’t define it. GrowLoop brings this within AI's reach:
📡 Disagreement isn’t noise
🎯 The rubric is the target
🧬 Criteria and cases co-evolve
#AIEval #LLMs #AIlignment

[AmapAI @Alibaba_AMAP]

IROS'26 3/3 [Diff-CAST]
✅ Diffusion-based Motion Priors
✅ No Mode Collapse (replaces GANs)
✅ Drift-free Sim2Real (SACC)
✅ Hardware-safe RL
Scaling quadruped skills with ease! 🐕
论文地址：https://t.co/LtcmEAX4MY
#IROS2026 #Robotics #RL

[AmapAI @Alibaba_AMAP]

IROS'26 1/3 [ABot-Explorer]
✅ Online Unified Exploration & Memory
✅ VLM-based Semantic Affordances
✅ SOTA Efficiency & Coverage
Mimicking human logic for Embodied AI. 🚀
📝https://t.co/lQeDuDSixq
💻https://t.co/lHQHNNDsm3
#IROS2026 #EmbodiedAI #VLM

[AmapAI @Alibaba_AMAP]

IROS'26 2/3 [AsyncShield]
✅ Fixes VLA Cloud Latency
✅ Deterministic Spatial Mapping
✅ Zero-shot Safety (CMDP)
Plug-and-play protection for real-world robotics! 🚀
📝https://t.co/P4gIEy6E4B
#IROS2026 #Robotics #VLA

[AmapAI @Alibaba_AMAP]

Testing 12 models reveals a gap: frontier LLMs excel at explicit tasks but fall short in implicit decision-making and proactive info-seeking. 📉🔍
The code will be open-sourced soon! Stay tuned. 🚀
#MapAgent #Benchmark #ImplicitNeeds #OpenSource

[AmapAI @Alibaba_AMAP]

Current map agent eval focuses on task completion but ignores user acceptance. MapSatisfyBench shifts focus to Decision Satisfaction by quantifying implicit needs from real interactions. 🗺️📍
📝https://t.co/g6nHyOrVZf
#MapAgent #DecisionSatisfaction #LLM

[AmapAI @Alibaba_AMAP]

@manishipsfast Cool! Keep us posted on how it goes. We're here if you hit any snags. 👊

[AmapAI @Alibaba_AMAP]

Introducing DreamX-World 1.0 — a general-purpose world model with 1 minute continuous generation, real-time interaction, precise camera control & multi-style support. Beta coming soon!
🌐https://t.co/U5NIvOi6rU
Github：https://t.co/zolKYnuWjT
#WorldModel #AIVideo

[AmapAI @Alibaba_AMAP]

@manishipsfast We highly recommend running our code on a GPU-enabled server or PC. You can still run it on a MacBook, but keep in mind that video generation will take considerably longer. If you need help when reproducing results on a MacBook, feel free to open a GitHub issue.

[AmapAI @Alibaba_AMAP]

Real-Time Control✔️Seamless viewpoint navigation through 4D space.
Long-Term Memory✔️One full orbit, everything remains intact.
Diverse Styles✔️Aesthetics have no standard; your prompt is the standard.
1min Generation✔️In the age of AI, let's create long-form video together.

[AmapAI @Alibaba_AMAP]

This is how we push the boundaries of real-world robot navigation. Next time Jensen Huang visits Beijing, we’d love to have him walk TuTu in person! 😎🧥🤖
@NVIDIARobotics @DrJimFan

[AmapAI @Alibaba_AMAP]

An incredible showcase of Embodied AI in the wild! 🐕⚡️
Meet TuTu (途途), navigating Beijing's chaotic morning rush hour.
By integrating a 3B VLA navigation model with the NVIDIA Jetson AGX Orin, we brought large-model reasoning straight to the edge.
#Robotics #Nvidia #VLA

[AmapAI @Alibaba_AMAP]

Why this matters for Edge AI:
🔹 Zero-Shot Wild: No remote control or HD maps. Pure vision-to-action reasoning.
🔹 AGX Orin Power: Dodging crowds, navigating narrow paths, and obeying traffic rules entirely on the edge.

[AmapAI @Alibaba_AMAP]

📝https://t.co/aubTcKz4CX
🌐https://t.co/9i0eUJ0DtU

[AmapAI @Alibaba_AMAP]

🚀 ABot-Earth 0.5👉🏻The world's first 3D native urban world model!
✨ Multi-modal: Supports both image & text inputs.
⚡️ Efficient: Generates 1 sq km high-fidelity 3D city scene in <10 mins.
💻 Accessible: Runs on a single consumer-grade GPU!

#AI #worldmodel

[AmapAI @Alibaba_AMAP]

📢AGenUI is now OPEN-SOURCE! Qwen & Amap release an end-to-cloud A2UI framework for iOS, Android & HarmonyOS.
High-performance, consistent, and extensible—built for the future of AI-native experiences.
https://t.co/rKbY6lr487
#OpenSource #AGenUI #Qwen #AmpaAI #A2UI #AgenticUI

[AmapAI @Alibaba_AMAP]

🏆 NeoVerse-ABot ranks #1 in the AGIBOT World Model Challenge! 🤖
Key Highlights:
✨ Visual Quality: 0.6246 (#1)
🎯 Action Following: 0.9651 (#1)
We’ve moved beyond "pretty videos" to "controllable results"—turning signals into precise physical interactions. 🚀
#AI #WorldModel

[AmapAI @Alibaba_AMAP]

Amap Skill Claw: Making AI skills persistent. Agents now evolve across sessions, devices & users. Just chat—it auto-extracts & verifies skills in the background. A seamless, self-evolving loop. 🔄✨
https://t.co/19elOJI8mj
#Amap #AI #agent #skill

[AmapAI @Alibaba_AMAP]

ABot-Claw: System-Level Embodied AI🌐🤖
Unifying hetero robots via a cognitive harness.

🔹Unified Scheduling
🧠Shared Memory
⚖️Hierarchical Control
🤝Social Alignment
Proven in fuzzy commands & cross-robot guidance. 🚀

🔗 https://t.co/svV05yHtZP
#AGI #AI #Robotics