SetupAI
繁
SourceFreshcollected in 14h

JD Open-Sources an Audiovisual World Model

Read original on Pandaily
#world-model#embodied-ai

JD.com offers open-source audiovisual world models for navigation and embodied-AI experiments.

30-Second TL;DR

What Changed

JoyAI-EchoWM and EchoWM-Flash are now open source

Why It Matters

Open sourcing could accelerate research into interactive world models and embodied agents. JD.com's planned domestic GPU cluster and Logistics Superbrain 3.0 also indicate a broader push to deploy AI across logistics.

What To Do Next

Download JoyAI-EchoWM and reproduce its WBench Navigation setup to evaluate audiovisual world-model performance on your own embodied-agent tasks.

Who should care:Researchers & Academics

Key Points

  • •JoyAI-EchoWM and EchoWM-Flash are now open source
  • •The models score approximately 81.6–81.7 on WBench Navigation
  • •They combine camera-intent control with native audio-video generation

Deep Insight

Background and context from public sources — not the original article. 9 sources cited.

Enhanced Key Takeaways

  • •The release serves as the cornerstone for JD.com's 2026 operational strategy, 'JoyAI: Leaping into the Physical World', aimed at anchoring foundation models to physical fulfillment and smart supply chains.
  • •JD.com paired the model announcement with logistics Superbrain 3.0, a decision engine capable of executing complex route optimization across hundreds of millions of packages within seconds.
  • •JD Cloud committed dedicated infrastructure to support the physical-world models, deploying a large-scale domestic GPU cluster to maintain compute and inference sovereignty.
  • •EchoWM enables real-time navigation through generative 3D virtual environments via active camera-pose directives and user intents rather than relying on passive video playback.
  • •The model launch builds on JD's sequential JoyAI rollouts earlier in 2026, including the instruction-tuned JoyAI-LLM Flash in March and the JoyAI-VL-Interaction streaming model in June.

Competitor Analysis

Model Accessibility
JD.com (JoyAI-EchoWM / EchoWM-Flash)
Open-source
World Labs (Atlas)
Proprietary / Closed API
LingBot-World
Open-source
Core Capabilities
JD.com (JoyAI-EchoWM / EchoWM-Flash)
Native audiovisual co-generation
World Labs (Atlas)
Generative 3D spatial intelligence
LingBot-World
Interactive world modeling
Control Mechanism
JD.com (JoyAI-EchoWM / EchoWM-Flash)
Unified camera-intent control
World Labs (Atlas)
3D spatial prompt directives
LingBot-World
Action/camera trajectory control
Target Benchmark
JD.com (JoyAI-EchoWM / EchoWM-Flash)
WBench Navigation (81.6–81.7)
World Labs (Atlas)
Custom spatial world evaluations
LingBot-World
General world model navigation baselines

Technical Deep Dive

  • Native Audiovisual Co-generation: Generates synchronized multi-channel audio tracks alongside visual frames in a unified synthesis step, avoiding decoupled post-production audio synthesis.
  • Unified Camera-Intent Control: Supports interactive scene exploration via explicit camera coordinate transformations combined with high-level user intent directives, replacing pre-scripted generation trajectories.
  • Spatial Navigation Benchmarks: Scores 81.6–81.7 on WBench Navigation, demonstrating consistent physical and geometric coherence across extended exploration sequences.
  • Dual-Model Portfolio Architecture: Includes full-scale JoyAI-EchoWM for maximum visual fidelity and EchoWM-Flash for low-latency interactive simulation, backed by domestic enterprise GPU acceleration.

Future ImplicationsAI analysis grounded in cited sources

Physical supply chain operations will increasingly integrate generative spatial simulation
JD's pairing of EchoWM with logistics Superbrain 3.0 indicates that world models are moving from passive media generation to active digital-twin simulation and routing control in automated fulfillment.
Open-source spatial world models will accelerate enterprise adoption over proprietary APIs
By releasing EchoWM openly with competitive navigation scores, JD provides developers with an accessible alternative to proprietary world models like World Labs' Atlas.

Timeline

2026-03
JD.com launches instruction-tuned foundation model JoyAI-LLM Flash
2026-06
JD.com open-sources JoyAI-VL-Interaction real-time vision-language stream model
2026-09
JD.com unveils logistics Superbrain 3.0 and domestic GPU cluster infrastructure expansion
2026-09
JD.com open-sources JoyAI-EchoWM and EchoWM-Flash interactive audiovisual world models at JDD
Read original article on Pandaily

Weekly AI Recap

Read this week's curated digest of top AI events →

PostLinkedIn

AI-curated news aggregator. All content rights belong to original publishers.
Original source: Pandaily ↗

This is a summary, not the original. Read the source, or get the weekly briefing.

The weekly digest

One email a week. Unsubscribe anytime.