Features
Direct Video Action (DVA) model architecture for generalist robot policies
Web-scale video pre-training on over 1 million videos
Post-training on just 1-10 hours of trajectory data
FutureVision video-predictive control for unseen task generalization
Long-context memory retention for in-context learning
Single-shot task execution from human demonstrations without retraining
Custom robot platform with 25kg rated (40kg peak) payload
Safety-rated brakes in every actuator
3 years continuous operation at rated payload
Returns processing automation for logistics
Bearing decanting for automotive assembly
Contico box breakdown and debris clearance in manufacturing
Full-stack hardware+software solution for heavy-duty factory automation
Vision-based safety rated through safety-rated vision
High component stiffness and linear response for precise control
Steerable model with emergent capabilities (π0.7)
Open-world generalization for mobile manipulation (π0.5)
Vision-Language-Action (VLA) architecture
Multi-Scale Embodied Memory (MEM) for tasks >10 minutes
Real-time action chunking for low-latency control
RL Token for efficient online RL from human videos
FAST action tokenizer for 5x faster training
Open-source model weights and code (π0)
Human-to-robot transfer in VLAs at scale
Step-by-step task reasoning with human feedback
Generalist policy for any robot, any task
Precise manipulation with few hours of RL data
