Tags / 14 articles
Every article tagged machine-learning, newest first.
The loop of evaluation and improvement inside most reinforcement learning methods, worked through on the cliff grid and traced from dynamic programming to AlphaGo Zero.
reinforcement-learningmachine-learning
Both look at a camera image, read an instruction and output robot moves. A VLA starts from an image-and-text model and learns the moves. A WAM starts from a video model and also learns the frames that come next. Eight drawings on what that changes.
world-action-modelsvlaroboticsmachine-learning
Four ways a policy-gradient update can go wrong on a robot task: blaming actions for the past, rewarding every attempt, overshooting on reused data, and groups that teach nothing. Reward-to-go, baselines, PPO clipping and group advantages, with their limits.
reinforcement-learningmachine-learningrobotics
A fixed robot dataset can mislead a learner in two different ways: a hole, where it asks about actions nobody tried, and mistakes, where it copies bad demonstrations. Stay close to the data for the first, and weight actions by the reward for the second.
reinforcement-learningmachine-learningrobotics
Why copying demonstrations that go left and right around a box can send a robot into the box, and what flow policies and action chunks change.
reinforcement-learningmachine-learningrobotics
Why I am learning reinforcement learning, and the six problems that drive its algorithms: delayed reward, exploration, waiting, scale, reusing data and feedback.
reinforcement-learningmachine-learning
One selected branch of RL, the value-based methods from the trial-and-error idea to DQN, where each later method fixes a problem an earlier one handles badly. Then Q-learning, which nudges one score toward the reward plus its best guess about the next cell.
reinforcement-learningmachine-learning
Why waiting for the final reward is slow, and how borrowing the next state's estimate lets you learn during the walk, with one worked update.
reinforcement-learningmachine-learning
When states share a few weights, one update can teach states you never visited, and it can also blur two states together and move the target you're chasing.
reinforcement-learningmachine-learning
A value target has to assume what you do next: on-policy learning assumes the moves you actually make, and off-policy learning assumes a different policy's moves.
reinforcement-learningmachine-learning
Function approximation, bootstrapping and off-policy data each make RL practical. Off-policy value learners such as deep Q-networks (DQN) and soft actor-critic (SAC) use all three, even though together they can let a prediction feed its own target with nothing to pull it back down.
reinforcement-learningmachine-learning
Rules of thumb for a reinforcement learning target: freeze, take a second estimate, or take the minimum, and why the combinations table is one reading, not a law.
reinforcement-learningmachine-learning
After part 6: direct motor drive, learning between drawings, feedback during a stroke, and a twin tested across days. The repeat floor is not a claim of one-millimeter control.
so-101machine-learningcontrol
Where a neural network fits in a robot controller. A network that learns only what the physics model gets wrong does the same job as a Kalman disturbance estimate, with more freedom, and this is where recent research is going.
so-101controlmachine-learning