Class definition
A policy is not a robot body; hardware numbers stay on class pages.
Demos, parameters and gaps
Company demonstrations are not benchmarks. Blog parameter counts remain press unless a paper or model card confirms them. Missing artefacts stay —; unpublished licences and backbone-licence conflicts remain explicit.
Genesis-Embodied-AI/genesis-world engine.GENE-26.5 and GR00T N1 identity
GENE-26.5 is not Galbot GraspVLA, LDA-1B or GroceryVLA. GR00T N1’s paper names Fourier GR-1 and thanks 1X; it does not name Galbot bodies or Genesis Eno.
Genesis AI — the lab
Genesis AI emerged from stealth on 2025-07-01 with a $105M seed co-led by Eclipse and Khosla Ventures. Participating names in contemporaneous coverage include Bpifrance, HSG, Eric Schmidt, and Xavier Niel.
CEO Zhou Xian (CMU robotics Ph.D., 2024); co-founder Théophile Gervet (ex-Mistral; press also names Skild AI). Stated goal: a universal robotics foundation model plus a horizontal platform. [P1]
Announced 2026-05-06 as the company’s VLA / “robot brain.” Company materials describe a human-scale five-finger hand and a tactile e-skin data-collection glove.
Claimed for the glove (company / internal testing):
- a 1:1:1 glove→human-hand→robot-hand mapping;
- “100× cheaper” than typical teleop rigs;
- up to 5× data-collection efficiency.
The company blog assigns Genesis Hand 1.0 “20 active, back-drivable degrees of freedom.” That figure is a hand claim, not a robot-body DoF, and is not copied into any hardware table. [P2] [P3]
Demos and the Eno body
Demos: 20-step meal prep with one-handed egg cracking, pipetting, wire harnessing, Rubik’s cube, piano and four-object one-handed grasp. The later Eno body remains announced, without a numeric column.
Policy models — what is actually public
Helix architecture figures are company claims, not paper results.
Skild S1 and GR00T versions
Skild S1’s in-context-learner post publishes neither parameters nor weights. GR00T here means the N1-2B paper checkpoint; N1.6 and N1.7 remain separate versions.
| Metric | GENE-26.5 | GR00T N1-2B | OpenVLA 7B | π0.5 | Skild S1 | Figure Helix |
|---|---|---|---|---|---|---|
| Identity | ||||||
| Kind# | company VLA / “robot brain” [P2] | open VLA; dual-system VLM + DiT flow matching [R1] | open VLA on Llama 2 + SigLIP/DINOv2 [R2] | VLA; PaliGemma backbone + action expert [R3] | company “omni-bodied” foundation model; S1 is an in-context learner [P4] | company VLA; System 2 VLM + System 1 visuomotor policy [P5] |
| Public date# | 2026-05-06 [P2] | arXiv 2025-03-18 [R1] | arXiv 2024-06-13; CoRL 2025 [R2] | arXiv 2025-04-22; CoRL 2025 Oral [R3] | S1 post August 2026 [P4] | 2025-02-20 [P5] |
| Paper# | — | arXiv:2503.14734 [R1] | arXiv:2406.09246 [R2] | arXiv:2504.16054 [R3] | — | — |
| Scale (as published) | ||||||
| Parameter count# | — | 2.2B total; 1.34B in the VLM [R1] | 7B [R2] | 2B VLM (PaliGemma) + 300M action expert [R3] | — | S2 7B VLM; S1 80M [P5] |
| Stated training data# | — | real-robot trajectories, human video, synthetic; Fourier GR-1 collect named [R1] | 970k real-robot episodes from Open X-Embodiment [R2] | web + heterogeneous robot data; experiment results only [R3] | — | ~500 hours teleop; hindsight language labels [P5] |
| Published latency / rate# | — | 16-action chunk 63.9 ms on L40, bf16 [R1] | — | — | — | S2 7–9 Hz; S1 200 Hz (company post) [P5] |
| Public artifacts | ||||||
| Weights# | — | released with Isaac-GR00T; NVIDIA Open Model License [M1] | HF openvla/openvla-7b; backbone is Llama 2 (Community License) [M2] [R2] | openpi checkpoints; not a product API [M3] | — | — |
| Code license# | — | Apache-2.0 [M1] | project and training code published; see repository [M2] | Apache-2.0 (Physical-Intelligence/openpi) [M3] | — | — |
| Public API# | — | — | — | — | — | — |
| Public benchmark table# | — | paper tables only [R1] | paper tables only [R2] | paper tables only [R3] | — | — |
| Stated embodiments in the source# | company demos on Genesis hardware; Eno announced later [P2] | Fourier GR-1 collect; 1X thanked for hardware [R1] | Open X-Embodiment mixture (many arms) [R2] | Physical Intelligence mobile-manipulation experiments [R3] | company copy: many morphologies; no public eval table [P4] | Figure humanoids (company post) [P5] |
OpenVLA inherits Llama 2’s Community License despite its card’s MIT label.
Helix 02, π0.5 and suites
Helix 02 remains a later release. π0.5’s parameter split comes from its paper, not π0’s blog. Suite identities live in the Benchmarks atlas; policy rates stay here.
Smaller open VLAs and papers without public weights
SmolVLA and SpatialVLA are later checkpoints or papers. OpenVLA-OFT fine-tunes OpenVLA 7B; it is not a new pretrain. Gemini Robotics publishes a paper and company blog, but no parameter count or public weights.
| Metric | SmolVLA | SpatialVLA | OpenVLA-OFT | Gemini Robotics |
|---|---|---|---|---|
| Kind# | compact VLA on LeRobot; SmolVLM-2 plus a flow-matching action expert [R7] | 3D spatial VLA with Ego3D positional encoding and Adaptive Action Grids; Paligemma2 backbone [R8] | fine-tuning recipe on OpenVLA 7B, not a new pretrain [R9] | Gemini Robotics VLA plus Gemini Robotics-ER, built on Gemini 2.0 [R10] |
| Parameter count# | 0.5B on the Hugging Face card; paper says less than 0.5 billion [M6] [R7] | 3.5B [R8] | uses the existing OpenVLA 7B weights [R9] | — |
| Stated training data# | fewer than 30k community episodes [R7] | 1.1 million real-robot episodes (Open X-Embodiment subset plus RH20T) [R8] | — | — |
| Named result# | trains on a single GPU; paper also says consumer-GPU or CPU deploy [R7] | — | LIBERO success 76.5% → 97.1%; 26× action-generation throughput vs the OpenVLA baseline in the paper [R9] | Embodied Reasoning Question Answer (ERQA) set of 400 questions [R10] |
| Weights# | HF lerobot/smolvla_base [M6] | — | fine-tune of openvla/openvla-7b; suite checkpoints now public, e.g. moojink/openvla-7b-oft-finetuned-libero-spatial [M2] [M15] | — |
| Paper# | arXiv:2506.01844 [R7] | arXiv:2501.15830 [R8] | arXiv:2502.19645 [R9] | arXiv:2503.20020 [R10] |
| Public API# | — | — | — | — |
Gemini Robotics has no model card. SpatialVLA’s project page establishes identity; no weight dump is copied here. OFT suite checkpoints remain fine-tunes, separate from OpenVLA 7B. See the documented LIBERO-Spatial pair.
Related EPIC / Galbot models (already catalogued)
These already have pins and a paper→body map on the research page. They are not new columns here. GroceryVLA has no public arXiv.
CoRL 2025 · arXiv:2505.03233. Synthetic-only grasping VLA; homepage product stack.
RSS 2026 · arXiv:2602.12215. 1.6B latent world-action model; real figures on Galbot G1 and Unitree G1.
preprint · arXiv:2607.06988. Test-time steering of a frozen world-action model from human video.
Product model. No public arXiv. Not given an invented paper id.
Prior open policies — still the public baseline
Earlier research policies stay separate from Galbot products and later releases.
FAST tokenizer
FAST is a DCT + BPE action tokenizer, not a standalone VLA; do not invent policy parameters. Physical Intelligence’s /blog/fast returned HTTP 404 when checked; the research page is cited instead.
| Metric | Octo | RT-2 | π0 (2024 blog) | RDT-1B | FAST |
|---|---|---|---|---|---|
| Kind# | open transformer generalist policy [R4] | VLA; vision-language model used as a robot policy [R5] | company VLA blog; PaliGemma + action expert shape later stated in π0.5 [P6] | open diffusion Transformer for bimanual manipulation [R6] | frequency-space action tokenizer (DCT + BPE); not a standalone VLA [R11] |
| Paper / post# | arXiv:2405.12213 [R4] | arXiv:2307.15818 [R5] | company blog (not a paper) [P6] | arXiv:2410.07864 [R6] | arXiv:2501.09747 [R11] |
| Parameter count# | released checkpoints 27M and 93M [R4] | — | — | 1B (plus a 170M ablation checkpoint) [R6] | — |
| Stated training data# | 800k trajectories from Open X-Embodiment [R4] | web-scale VLM pretraining + robot data (paper) [R5] | — | 1M+ multi-robot episodes (card / paper) [R6] | FAST+ trained on 1 million trajectories; π0-FAST trains on 10k hours [R11] |
| Named result# | — | — | — | — | paper claims up to 5× faster training than diffusion VLAs when FAST is combined with π0; that is a training-time claim, not a policy parameter count [R11] |
| Weights# | public checkpoints on the project page [M4] | — | later openpi checkpoints belong to the π0.5 column above [P6] | HF robotics-diffusion-transformer/rdt-1b; MIT [M5] | research page; not a policy checkpoint [M9] |
| Public API# | — | — | — | — | — |
Open X-Embodiment (arXiv:2310.08864) is a dataset, not a policy.
RT-2, π0 and RDT-1B
RT-2 released no weights with its paper. π0’s blog does not establish π0.5’s parameter split. RDT-1B is Tsinghua TSAIL research, not Galbot or Genesis. Current openpi checkpoint names remain separate.
Current open releases — later versions, not replacements in place
These columns are later official releases or distinct openpi model IDs. They do not overwrite the N1-2B or π0.5 paper columns above. A GitHub release tag is not a success rate. Empty cells stay empty.
| Metric | GR00T N1.6 | GR00T N1.7 | π0 (openpi) | π0-FAST |
|---|---|---|---|---|
| Version status# | replaced by N1.7 in the current main branch; n1d6 tree kept [M10] [M13] | current GA on the Isaac-GR00T default branch [M10] | base flow-matching VLA still published beside π0.5 [M3] | autoregressive VLA using the FAST tokenizer; still published beside π0.5 [M3] [R11] |
| Public date# | GitHub n1.6-release 2026-04-15; patch n1.6.1-release 2026-04-23 [M13] | GitHub n1.7-release 2026-04-18 [M13] | — | — |
| Kind# | open VLA; Eagle VLM + flow-matching DiT (N1.6 path) [M10] [M12] | open VLA; Cosmos-Reason2-2B / Qwen3-VL backbone + flow-matching DiT; relative EEF action space [M10] | flow-based VLA [M3] [P6] | autoregressive VLA on FAST (DCT + BPE) [M3] [R11] |
| Parameter count# | 3B in the model id GR00T-N1.6-3B [M12] | 3B in the model id GR00T-N1.7-3B [M11] | — | — |
| Weights# | HF nvidia/GR00T-N1.6-3B [M12] | HF nvidia/GR00T-N1.7-3B; loads gated nvidia/Cosmos-Reason2-2B [M11] [M10] | gs://openpi-assets/checkpoints/pi0_base [M3] | gs://openpi-assets/checkpoints/pi0_fast_base [M3] |
| Checkpoint licence# | NVIDIA License on the card (non-commercial PDF linked) [M12] | NVIDIA Open Model License Agreement [M11] | see openpi assets; not a product API [M3] | see openpi assets; not a product API [M3] |
| Code licence# | Apache-2.0 [M1] | Apache-2.0 [M1] [M10] | Apache-2.0 [M3] | Apache-2.0 [M3] |
| Public benchmark table# | — | README names RoboCasa, SimplerEnv, and real G1 evals; no numeric table is copied here [M10] | — | — |
N1-2B paper figures stay separate. The README’s Apache-2.0 claim covers code; N1.7 weights use the NVIDIA Open Model License.
Parameter count and π0.5 weights
Do not transfer N1’s 2.2B count to N1.6/N1.7. π0.5 base weights remain gs://openpi-assets/checkpoints/pi05_base in the family column.
Documented evaluation pair — not a ranking
The pinned recipe defines the model, suite and checkpoint protocol. It does not reproduce the paper’s four-suite averages. Unknown fields are not zero; openpi’s π0.5 LIBERO results use a different protocol.
Result provenance
Paper-reported, upstream-reproduced and site-run results stay distinct; other models’ rates are not comparable here.
| Field | OpenVLA-OFT × LIBERO-Spatial |
|---|---|
| Result kind# | pinned runnable recipe; paper protocol is separate; site-reproduced: no [M15] [M18] |
| Suite / subset# | LIBERO / libero_spatial on sim-manipulation-eval. Suite pin 8f1084e3132a39270c3a13ebe37270a43ece2a01 (master HEAD). Not a tagged suite release. [R5] [M19] |
| Split# | no distinct train/eval split; eval is the 10 Spatial tasks with DEFAULT init states [M15] [M18] |
| Code / suite pins# | moojink/openvla-oft@e4287e9 (main); Lifelong-Robot-Learning/LIBERO@8f1084e (master). Current HEADs, not proof of paper scores. [M18] [M19] |
| Embodiment / action space# | LIBERO sim (Franka in the suite); continuous L1-regression; 8-action chunk; gripper normalize then invert; unnorm_key=libero_spatial_no_noops [M15] [M18] |
| Preprocess / runtime# | --center_crop True; 2 images + proprio; 10 wait steps; max 220 steps; env image 256. Paper runtime: Python 3.10.14, PyTorch 2.2.0, custom transformers v4.40.1, A100. [M15] [M18] |
| Checkpoint# | moojink/openvla-7b-oft-finetuned-libero-spatial HF 6d0231af0e48c5985f1ff86908f4674b84bc049b (2025-06-17). Weight LFS SHA-256s in worlds/models.json. [M20] |
| Licences# | code MIT; card MIT; inherited OpenVLA / Llama 2 terms still apply to the backbone [M15] [R2] |
| Training data# | openvla/modified_libero_rlds libero_spatial_no_noops; LoRA r=32 on openvla/openvla-7b [M15] |
| Episodes / seeds# | script default 10 tasks × 50 episodes (500); --seed default 7. Paper used three unpublished seeds. [M15] [M18] |
| Named metric# | Spatial-only success unpublished. Paper four-suite average of independently trained OFT policies: 97.1%; combined policy 96.8%. Those paper figures are not this pin. [R9] [M15] |
| Hardware / date# | paper: NVIDIA A100. Pin date 2026-09-08. Eval date of the paper numbers unpublished. [M15] |
| Upstream command# | python experiments/robot/libero/run_libero_eval.py --pretrained_checkpoint moojink/openvla-7b-oft-finetuned-libero-spatial --task_suite_name libero_spatial [M15] [M18] |
openpi’s upstream-reproduced π0.5 @ 30k LIBERO results: Spatial 98.8, Object 98.2, Goal 98.0, LIBERO-10 92.4, average 96.85; checkpoint gs://openpi-assets/checkpoints/pi05_libero/. [M14]
Comparability
Episodes, seeds and hardware are unpublished. These upstream-reproduced results are neither site-reproduced nor comparable with this column.
LeRobot inventory — code listing is not a weight dump
LeRobot’s README lists policies it implements. A row here is that listing plus any Hub checkpoint seen in this checkout. Framework support is not a pretrained-weight proof and not a claim those models are SOTA.
| Metric | VLAs listed | World models listed | Reward models listed |
|---|---|---|---|
| Names in the README# | Pi0, Pi0Fast, Pi0.5, GR00T N1.7, SmolVLA, XVLA, EO-1, MolmoAct2, WALL-OSS, EVO1 [M16] | VLA-JEPA, LingBot-VA, FastWAM [M16] | SARM, TOPReward, Robometer [M16] |
| Public Hub weights seen# | smolvla_base, pi0_base, pi05_base, pi0fast-base, xvla-base, eo1-base, wall-oss-0.5, MolmoAct2-SO100_101-LeRobot; EVO1 none found [M16] [M6] | VLA-JEPA-Pretrain, VLA-JEPA-LIBERO, lingbot_va_base, fastwam_libero_uncond_2cam224 [M17] | — (no lerobot/sarm, topreward, or robometer model found) [M16] |
| Kind# | action-generating VLAs [M16] | predicted video/state or world-action heads; not a physical success proof [M16] | reward estimation; not action generation [M16] |
| Named eval hook# | README shows lerobot-eval on LIBERO with lerobot/pi0_libero_finetuned as an example command, not a result table [M16] | — | — |
A visually plausible world-model rollout is not a successful physical execution.
Galbot / EPIC models and version records
Galbot / EPIC models stay on the research cards above unless a compatible public artefact is published. Machine-readable version records live in worlds/models.json; this page does not fetch that file.