Vision-Language-Action Models for Robotics Quiz

Answers and explanations appear as you take the quiz.

1. What three types of input/output define a Vision-Language-Action (VLA) model?
   • Camera images, a language instruction, and robot actions
   • Audio, text, and images
   • Sensor logs, GPS, and battery state
   • Code, compiler output, and test results
2. How did Google DeepMind's RT-2 represent robot actions so a transformer could generate them?
   • As continuous floating-point vectors only
   • As discretized action tokens, similar to text tokens
   • As ROS message XML
   • As reward values for reinforcement learning
3. What does 'emergent transfer' mean in the context of VLAs?
   • The robot overheats and shuts down
   • Two robots swap their software
   • Web-scale knowledge from pretraining helps the robot handle novel objects and instructions
   • Actions are computed without any neural network
4. What is a key advantage of pi-0's diffusion/flow-matching action head compared to single discrete action tokens?
   • It removes the need for any cameras
   • It only works in simulation
   • It eliminates the language instruction
   • It generates smooth, high-frequency action chunks for more dexterous control
5. Why does the Open X-Embodiment dataset matter for VLAs?
   • It pools robot data across many robot types to enable cross-embodiment generalization
   • It is a benchmark only for image captioning
   • It stores only one robot's data
   • It is a language-only text corpus
6. Which data sources did DeepMind combine when co-fine-tuning RT-2?
   • Only a fixed list of hand-written joint commands
   • Robot demonstration data and web data with a pretrained vision-language model
   • Only audio recordings without robot observations
   • Only random action tokens with no examples
7. What kinds of commands can an RT-2-style action representation encode?
   • Only natural-language explanations with no robot controls
   • A fixed video caption for each camera frame
   • Discretized robot motion and gripper commands represented as tokens
   • A robot’s hardware serial number
8. How broad was the RT-2 evaluation of emergent capabilities reported by Google DeepMind?
   • Exactly 12 trials
   • One trial for every robot type worldwide
   • No physical robot trials
   • More than 6,000 robotic trials