Deep Robotic Learning
summary
Argues end‑to‑end sensorimotor learning beats hand‑engineered pipelines by closing perception and control, shows concrete algorithms and tradeoffs: DAgger fixes distribution mismatch by querying labels from the policy’s state distribution, PLATO uses MPC with a KL penalty to collect safe training data, and guided policy search trains vision policies from state‑based trajectories. Demonstrations include a forest‑flying quadrotor, simulated humanoid running, grasping learned on many robots (2.5 hours total) and model‑based manipulation learned in ~3 minutes of data.
AI-written summary · about
chapters (9)
- 0:00Introduction to robotics
- 1:18Perception and action loops
- 8:11Challenges in robotic control
- 11:51Imitation learning
- 30:27Optimal control foundations
- 35:56Safe policy learning
- 46:07Trajectory optimization
- 59:11Reinforcement learning
- 1:19:04Data efficiency and future
related talks
18mRobotics: why now?
28mPreference Learning from Minimal Human Feedback for Interactive
17mBeyond Automation: AI and the Evolution of Robotic Intelligence
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
35mIndustrial Applications of Reinforcement Learning
49mML: Teaching machine to learn from its own mistakes (...)
05mPractical insights into deep reinforcement learning
32mThe Road to General AI