PlayWorld: Learning Robot World Models from Autonomous Play
robot-playworld.github.io · 597 words · saved by 1 readers
PlayWorld: Learning Robot World Models from Autonomous Play
We introduce PlayWorld, a scalable framework for training high-fidelity Action-Conditioned Robot World Models using Autonomous Play. Instead of relying on success-biased human demonstrations, it continuously gathers diverse, contact-rich interactions and failure cases that better match downstream policy behaviors. The resulting model enables accurate dynamics prediction, reliable policy evaluation, and effective world-model-based RL fine-tuning that improves real-world performance. Collecting Play Data 🧱 Blocks 🥕 Carrot 👕 Towel System overview for autonomous play-data collection.…
saved by
related reading
- [2603.08546] Interactive World Simulator for Robot Policy Training and Evaluationarxiv.org
- Precise Manipulation with Efficient Online RLpi.website
- World Models | Rohit Bandarurohitbandaru.github.io
- State of Robot Learning, December 2025vedder.io
- A Steerable Model with Emergent Capabilitiespi.website
- Robotic World Model: A Neural Network Simulator for Robust Policy Optimization in Roboticsarxiv.org
- 1X World Model | From Video to Action: A New Way Robots Learn1x.tech
- Generalist - GEN-0 / Embodied Foundation Models That Scale with Physical Interactiongeneralistai.com
- Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models | NVIDIA Technical Blogdeveloper.nvidia.com
- Learning Latent Plans from Playlearning-from-play.github.io
- A Unified Model for Motion-Conditioned Robot Co-designtransformer-transformer.github.io
- How Claude Performs on Robotics Tasks \ Anthropicanthropic.com