Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·28d ago·5 min read
Pelican-Sim 1.0: A General World Model Simulator for Embodied Intelligence
TL;DR
Pelican-Sim 1.0 is a versatile world model simulator designed for embodied intelligence, achieving significant improvements in action controllability and video quality across various robotic tasks.
✦ Why It Matters
Explore Pelican-Sim 1.0 to enhance your robotic simulations and improve decision-making processes today.
Key Takeaways
- Utilize Pelican-Sim 1.0 for diverse robotic applications.
- Leverage the unified action representation for cross-platform compatibility.
- Implement action-visual injection to enhance controllability.
- Adopt the sparse MoE architecture for better dynamic management.
Full Summary
Pelican-Sim 1.0 introduces a general world model simulator that predicts future observations based on visual context and robot actions, enhancing learning and decision-making for embodied agents. Key innovations include a unified 28-dimensional action representation, which allows compatibility across different robotic platforms, and action-visual injection that improves controllability by linking actions to visual outputs.
The model employs a sparse mixture-of-experts (MoE) architecture to manage diverse dynamics while minimizing conflicts between action modalities. Additionally, it features efficient rollout generation, achieving a 5.67-fold speedup in simulation time.
Trained on nearly one million trajectories, Pelican-Sim demonstrates substantial gains in performance metrics, such as a PSNR improvement of 4.636 on AgiBotWorld Beta. The simulator's effectiveness is validated through four applications on RoboTwin, significantly increasing policy success rates and demonstrating high correlation in policy evaluation.