Primary sources for the 2026 VLA frontier¶
The October 18 source update supplements the original snapshot with current access and pinned artifact details.
Read these alongside Lectures 25–30, using the dated frontier register to distinguish announcements, papers, and executable releases. Checked 2026-10-13. Sources support only the nearby claims in the lectures; the course's calculations and fixture designs are original teaching exercises, not measured results from these systems.
Physical Intelligence — read architecture, then artifacts¶
- π0 paper and version history — original architecture and dated submissions.
- π0 code/weight release announcement — distinguishes publication from release.
- π0.5 paper — heterogeneous supervision and generalization experiment framing.
- Knowledge-insulation paper — the training-dynamics question behind continuous action experts.
- π*0.6 paper, full text, MEM paper, and official research index — distinguish announcement dates from arXiv submission dates.
- π0.7 announcement, paper/version record, and full text — arXiv access succeeded after publisher-hosted PDF retrieval returned 403.
- openpi code and checkpoint inventory — choose a reproducible baseline here; inspect code terms and Gemma terms separately.
NVIDIA — read version history before installation¶
- GR00T N1 paper — historical architecture and evaluation.
- N1.7 early-access announcement, release tags, and commit history — distinguish EA and GA events.
- N1.7 technical article — dated general-release context.
- Isaac-GR00T repository, hardware guidance, and FAQ — pin a revision before applying recipes.
- N1.7 3B model card — observed version/license labeling differs from the current repository; preserve the discrepancy in a model manifest.
Google DeepMind — identify which model is accessible¶
- Initial Robotics/ER announcement.
- On-Device announcement.
- Robotics 1.5 / ER 1.5 announcement.
- ER 1.6 announcement.
- Robotics 2 family announcement.
- ER 2 model card and On-Device 2 model card — keep API access, partner access, and local weights separate.
Open policies and representation learning¶
| Study | Paper/date evidence | Author artifacts to inspect |
|---|---|---|
| OpenVLA-OFT | Paper | Code and evaluation guides |
| SmolVLA | Announcement | Base checkpoint, versioned LeRobot documentation |
| RDT | Paper | Author repository |
| MolmoAct 2 | Announcement and updates, paper | Code and model inventory, model collection |
| LingBot-VLA | 1.0 release history, 2.0 paper | 2.0 code, 2.0 6B card |
| Qwen-VLA | Paper and version history | Executable artifact availability not established by this review |
| VLAct | Paper v2, 2026-10-06 | Code/model zoo; read artifact-versus-paper configuration notes |
Reasoning, world models, and evaluation¶
- DreamZero paper and code/checkpoints — joint world/action prediction with an actual deployment path.
- Continuous Reasoning for VLA — a dated research study of intermediate representations; code/weight availability was not established here.
- In-Context VLA — grounded context and tool acquisition; paper verified, executable artifacts not established.
- VLAQuantBench — a recent closed-loop quantization study; follow its linked artifacts when reproducing it.
- RoboCasa365 official leaderboard — benchmark-specific October 13 snapshot; inspect protocol and individual artifact links before comparing entries.
- Figure Helix 02 and Helix 2.5 — whole-body hierarchy and a September 2026 generalization study; public weights were not established.
Reading deliverable¶
Choose one paper, one repository, and one model card for the same candidate. Write a three-column audit: reported claim, available artifact, and your executable test. Record missing information explicitly. The objective is to know what you can reproduce and what remains an author-reported result.