Manipulation & Learned Policies
Manipulation is a robot handling objects with its arms and hands, and a learned policy maps what the robot senses to its next action.
- Behavior Cloning Foundations
Covariate shift makes naive imitation drift, and DAgger is the standard fix.
- Action Chunking (ACT and ALOHA)
ACT predicts short action sequences and smooths them with temporal ensembling.
- Diffusion Policy
Visuomotor control as conditional denoising over action sequences, with receding-horizon execution.
- Vision-Language-Action Models
RT-1, RT-2, Octo and OpenVLA, and what web-scale pretraining adds to robot control.
- The Pi Line
From pi0 to pi0.7, with flow-matching action experts and FAST tokenization.
- Other Generalist Policies
How to read the claims for Gemini Robotics, GR00T, Helix and GO-2.
- Comparison Matrix
Every major policy compared on eight architectural axes.
- Hierarchical Approaches
From SayCan and Code as Policies to the hierarchies in Helix and pi0.5.
- RL Fine-Tuning of Policies
Improving a pretrained policy from its own online experience.
- Real-Time Execution
The latency budgets that decide whether the control loop closes.
- Cross-Embodiment Transfer
How one policy learns from many robot bodies and from humans.
- Knowledge Insulation
Discrete tokens train the backbone while the action expert trains behind a stop-gradient.
- Robot Learning Curriculum for ML Engineers
A dependency-ordered route from supervised learning to real robot policies.
- Action Spaces for Robot Learning
What each action representation gives the learner and asks of the controller.
- Foundation Models for Robotics
What foundation means in robotics, and what adaptation still costs.