Excited to share that I’ve joined @AnthropicAI! Looking forward to learning from an incredible team and working on exciting problems at the frontier of AI.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Hi all, I'm excited to announce a completed project - the ColosseumV2 (accepted to RA-L/ICRA)! The ColosseumV2 is a GPU-parallelized benchmark designed to evaluate the generalization and performance of modern VLA models. It's a follow up to the original 2024 Colosseum paper. What does V2 do differently? It's much faster, has more tasks, perturbations, and multiple morphologies and it's built on ManiSkill which means it can run headless the codebase is flexible. The benchmark contains 28 tasks, with 13 task categories, 2 robot morphologies, and 19 perturbations. Inspired by the VLA training recipe itself, we have decomposed generalization into three axes: vision, language, and action. Pi0.5, MolmoAct2, and ACT are all evaluated for reference.
Website: colosseum-v2.github.io/
Arxiv: arxiv.org/abs/2605.27759
Code: github.com/jstmn/ColosseumV2…
Examples of original (black rectangles) and synthesized (red rectangles) wrist-camera images from both arms using D-CODA on the simulated and real-world lift-ball tasks.
The original states are augmented every k time steps, resulting in out-of-distribution states. The non-augmented states remain in-distribution and guide the behavior cloning policies to complete the tasks.
To generate action labels (perturbed camera poses), we decompose a bimanual manipulation task into contactless and contact-rich states.
- Contactless states: sample random perturbations.
- Contact-rich states: apply identical constraint-optimized perturbations to both arms.
The diffusion model is an iterative denoiser that learns to map source wrist-camera images to target wrist-camera images, conditioned on pose transformations.
D-CODA is an offline data augmentation method for eye-in-hand bimanual imitation learning. Given a dataset, we train a diffusion model to synthesize novel view-consistent wrist camera images for both arms while generating action labels, producing an augmented dataset.
Learning-based systems still require large amounts of data for effective generalization. Our CoRL 2025 paper, D-CODA, expands data without simulators or environment interactions by generating novel view-consistent wrist-camera images and action labels for bimanual robots.
🧵👇
Examples of original (black rectangles) and synthesized (red rectangles) wrist-camera images from both arms using D-CODA on the simulated and real-world lift-ball tasks.
The original states are augmented every k time steps, resulting in out-of-distribution states. The non-augmented states remain in-distribution and guide the behavior cloning policies to complete the tasks.
To generate action labels (perturbed camera poses), we decompose a bimanual manipulation task into contactless and contact-rich states.
- Contactless states: sample random perturbations.
- Contact-rich states: apply identical constraint-optimized perturbations to both arms.