Tags / 8 articles

robotics

Every article tagged robotics, newest first.

October 23, 202625 min read

Do you need a human body to use human data?

Figure AI says a humanoid can learn from human video because it is shaped like us. About 30 papers with measured results say human video also trains grippers and wheeled robots, that mismatch breaks transfer, and that the body decides what a robot can reach. A companion to Three robot bets, with a pinned blackboard.

October 23, 202621 min read

Three robot bets: where does the generalization ceiling come from?

Figure, Generalist and Physical Intelligence each bet on a different source of robot generalization: the body, the data or the mix. A long read with a pinned blackboard.

robotics robot-learning foundation-models humanoids

October 23, 202614 min read

World action models vs VLAs: predict the video, or just the moves

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-models vla robotics machine-learning

Reinforcement learning, one problem at a time · Part 6 of 8October 21, 202613 min read

Policy gradients: fair credit and careful steps

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-learning machine-learning robotics

Reinforcement learning, one problem at a time · Part 7 of 8October 21, 202616 min read

Learning from a fixed dataset: missing actions and bad demonstrations

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-learning machine-learning robotics

Reinforcement learning, one problem at a time · Part 8 of 8October 21, 202614 min read

Mode averaging: two good routes, one bad average

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-learning machine-learning robotics

October 8, 20267 min read

Cycle time and reliability: shorter decisions and fewer task failures.

A few predictions about robotics

If intelligence keeps getting cheaper, the robotics industry may start to look more like a combination of AI infrastructure, industrial services, and manufacturing.

robotics industry ai

September 28, 202619 min read

Diagram of OpenVLA: the input image goes through DinoV2 and SigLIP encoders and an MLP projector, the instruction goes through the Llama tokenizer, Llama 2 7B processes both, and an action de-tokenizer outputs a 7D robot action: change in position, rotation and gripper.

Why I am learning robotics

The most impressive robots of the last decade stayed in videos because they could not generalize. Language models changed that, and robotics now looks like language models in 2019.

robotics industry

← All tags · All writing