Blog

Building a Robot Learning Pipeline with SO-101

Part II · VLA post-training with GR00T N1.7

Part II: adapting GR00T N1.7 to the carrot-to-box dataset with a full-scope LoRA interface.

VLA Post-Training

Part I · The LoRA family on GR00T N1.7

Part I: a controlled comparison of LoRA, QLoRA, and DoRA for post-training GR00T N1.7 on LIBERO-10 Long.

VLA Robot Learning

Part II · Data, embodiments, and open generalist policies

Part II: why multi-robot data is difficult to pool, and how open generalist policies adapt across embodiments.

VLA Robot Learning

Part I · From behavior cloning to generative action chunks

Part I: the policy foundations beneath VLAs—from behavior cloning to action chunks, diffusion, and flow matching.

Robot Kinematics

Part II · Forward and inverse kinematics

Part II of a technical note on robot kinematics, covering forward and inverse kinematics for serial robot arms.

Robot Kinematics

Part I · Coordinate frames, rotation, and rigid transforms

Part I of a technical note on robot kinematics, covering coordinate frames, rotation, and rigid transforms.

Building a Robot Learning Pipeline with SO-101

Part I · Hardware bring-up, teleoperation, and data collection

A practical build note on using SO-101 as a hardware platform for teleoperation, camera bring-up, and early dataset collection.

Implicit Neural Representations

Part III · Making radiance fields fast

Part III: faster radiance fields through hash encodings, tensor factorization, and explicit Gaussian primitives.

Implicit Neural Representations

Part II · Neural radiance fields: from camera rays to pixels

Part II: a neural radiance field turns calibrated images into a continuous scene that can render new views.

Implicit Neural Representations

Part I · From discrete samples to continuous fields

Part I: represent an image, shape, or scene as a continuous function that can be queried at any coordinate.

Multi-View Geometry

Part III · Computing the fundamental matrix from point correspondences

Part III turns the epipolar constraint into an algorithm: estimating a fundamental matrix from noisy point matches.

Multi-View Geometry

Part II · Epipolar geometry and the fundamental matrix

Part II explains why a point in one image can correspond only to an epipolar line in another, and how the fundamental matrix expresses that constraint.

Multi-View Geometry

Part I · From 3D points to image pixels

Part I builds the camera model from an intuitive idea of projection to the geometry that maps a 3D world point to a 2D pixel.

Parameter-Efficient Fine-Tuning

Part I · LoRA, QLoRA, and DoRA

Part I: LoRA, QLoRA, and DoRA—three ways to adapt a large frozen model without fully fine-tuning it.

Rethink Human Learning Through the Lens of Machine Learning

Machine learning offers a surprisingly useful vocabulary for thinking more systematically about how humans learn, adapt, and make decisions.

Labor Day at Riverside House

A quiet Labor Day weekend by the river: a house, board games, and an unhurried afternoon outside.