Model zoo

F.1 The table

F.2 How to read the zoo

The table is laid out so that consecutive rows roughly correspond to the order in which the models are introduced in the book. Rows 1–4 (PaLM-E through RT-2-X) are the Google DeepMind lineage covered in Chapters 11–12. Rows 5–6 (OpenVLA, Octo) are the open-source generalist policies treated in Chapter 12 and revisited in Chapter 16. Rows 7–13 (RDT-1B through RoboMamba) are the continuous- action and efficient-VLA family of Chapter 13 and §16.x. Rows 14–17 (Helix through ρα) are the dual-system models of Chapter 14. Rows 18–24 (Embodied-R1 through OpenDriveVLA) are the specialized and adjacent models — long-horizon, 3D, cross-embodiment, driving — covered in Chapter 15.

Three caveats on the parameter-count column. The number reported is the total parameter count of the deployed model unless otherwise noted; for dual-system models the column reports the sum of the slow (VLM) and fast (sensorimotor) components and the breakdown is given in the corresponding chapter. For models that ship in multiple sizes (RT-2 at 5B and 55B, Octo at 27M and 93M, RoboBrain2.0 at 7B and 32B), the column reports the range; the chapter treatment uses the size the authors emphasize. For models that report only an active parameter count or a sparse-expert total, the chapter body explains the convention used; the zoo cell is the total.

F.3 What the zoo deliberately omits

A few well-known systems are absent from the table by design. Standalone vision-language models (Llama-3 family, Qwen-VL, PaLI-X) are excluded — they are backbones for VLAs in this table, not VLAs themselves. Pure robotics policies without language conditioning (Diffusion Policy, ACT) are excluded — they appear in Chapter 10 as architectures rather than as foundation models, and the distinction matters: a Diffusion Policy trained on one task is not a foundation model, but it is part of the lineage that produced the action heads of π0 and RDT-1B. Pure planners with LLM “brains” (SayCan, Code as Policies, Inner Monologue) are excluded — they emit symbolic plans rather than low-level actions, and so technically belong to the lineage of Chapter 4 rather than the zoo of Part 4.

The boundary between “foundation action model” and “large policy” is not crisp; the inclusion criterion used here is that the model (i) ingests both vision and language and emits actions, (ii) has been pretrained or co-trained on a corpus that includes general web-scale content rather than only robotics data, and (iii) is presented by its authors as a step toward generality rather than as a single-task demonstration. This is the same criterion the surveys cited in Chapter 1 use, and disagreement at the boundary (Octo vs. RT-2; LEO vs. PaLM-E) is the kind of disagreement that the introductory chapter of each survey spends a section on.

F.4 Cross-references

A small number of cross-cutting facts may be useful when comparing rows.

The single largest parameter count in the table belongs to PaLM-E (up to 562B); the single smallest is SmolVLA (450M). The ratio is about 1,250×, which is the dynamic range the field currently operates over.

The four models that use a flow-matching action head — π0, SimVLA, Xiaomi-Robotics-0, SmolVLA — share the architectural lineage laid out in Chapter 10 and revisited in Chapter 13. The two that use a diffusion-transformer head — RDT-1B and TinyVLA — share a different lineage; the contrast is the substance of §10.5.

The two dual-system models — Helix and GR00T N1 — share the System- 1 / System-2 architectural pattern but disagree on the latency budget assigned to each system; Chapter 14’s §14.4 has the breakdown.

The open-source weights, in order of permissiveness, are: OpenVLA (Apache 2.0), Octo (MIT), RDT-1B (Apache 2.0), SmolVLA (Apache 2.0), TinyVLA (research license), GR00T N1 (NVIDIA research license), π0 (Physical Intelligence research license). RT-1, RT-2, PaLM-E, and the Helix family are not openly weighted. This matters for Chapter 16; the recipes there assume you have a checkpoint, and the open-weight column of the table is the candidate list.

For the full bibliographic entries of every primary reference in this table, see Appendix E.2. For the lineage that produced this zoo — CLIP → RT-1 → RT-2 → OpenVLA → π0 → Helix / GR00T N1 — see Chapters 11 through 14. For the open scientific problems that remain across the zoo, see Chapter 18.