VLA frontier register — checked 2026-10-13¶
This is the evidence map for Lectures 25–30. Use it to choose an executable study and to distinguish a research claim from an available artifact. All rows were checked on 2026-10-13; the cutoff is that date, not 2026-10-31. This selected register is not an exhaustive literature census or a universal model ranking.
Later review: the October 18 update adds the Pi API beta, pins inspected source interfaces, clarifies GR00T code/weight terms, and records the live RoboCasa page's October 16 update. The historical rows below retain their original review scope.
Dates below identify their event: paper submission, announcement, code release, or revision. A moving repository is an access observation, not a reproducibility pin. Before executing a lab, resolve and record its exact code commit, model revision/hash, dataset revision, configuration, normalizer, runtime, and applicable artifact licenses. No full-model benchmark or hardware reproduction was performed for this register.
Physical Intelligence and NVIDIA¶
| Artifact | Dated event and primary evidence | Architecture/action interface | Access and license evidence | Evaluation boundary; feasible reproduction |
|---|---|---|---|---|
| π0 | Paper: 2024-10-31; v4 2026-01-08. Code/weight announcement: 2025-02-04 | VLM-conditioned continuous flow action chunks | openpi provides artifacts; code license and Gemma terms require separate accounting | Run an appropriate released checkpoint or adapt it; public artifacts do not reconstruct every private training/evaluation episode |
| π0.5 | Paper: 2025-04-22 | Heterogeneous co-training and semantic/action supervision | Public base and adapted checkpoints in openpi; inspect checkpoint-specific terms | Evaluate released task paths; a LIBERO-adapted policy does not reproduce new-home experiments |
| Knowledge insulation | Paper: 2025-05-29 | Training interaction between pretrained representations and continuous action expert | Research method; implementation coverage depends on selected openpi version | Test gradient-routing/training hypotheses only where the implementation exposes the relevant controls |
| π*0.6 / experience learning | Official index: announcement 2025-11-17; paper: November 18, v2 November 19 | RECAP advantage-conditioned improvement | Full arXiv HTML accessible; corresponding downloadable weights not verified | Audit real-robot task success/throughput protocol; do not label a π0.5 run a π*0.6 reproduction |
| Multi-scale embodied memory | Official index: announcement 2026-03-03; paper: March 4, v2 March 8 | Video short-term and text long-term memory | arXiv paper accessible; full released policy artifacts not verified | Reproduce the course's memory fixture; distinguish it from author experiments |
| Efficient online RL for precise manipulation | Official index: 2026-03-19 | Precise-task policy improvement | Dated official summary; executable release not verified | Research pointer only in this review |
| π0.7 | Announcement/paper: 2026-04-16; v2 April 24 | Steerable generalist with multimodal context | Full arXiv HTML accessible; public π0.7 checkpoint not verified | Distinguish coached, high-level-policy, and direct execution conditions; use an explicitly different public baseline for labs |
| GR00T N1 | Paper: 2025-03-18; v2 March 27 | Coupled vision-language and generative action components | Historical repository releases; terms depend on version | Follow historical configuration when reproducing N1, rather than current N1.7 defaults |
| GR00T N1.7 | EA announcement: 2026-04-17; EA tag April 18; GA article/commit history: 2026-07-07 | Cosmos-Reason2-2B backbone; continuous action chunks; embodiment-specific contracts | Repository and 3B card provide code/weights but differ on licensing/EA labeling; see note below | Supported task evaluation and adaptation are possible; benchmark-specific checkpoints, camera inputs, and hardware setup must match |
Google DeepMind¶
| Artifact | Dated event and primary evidence | Architecture/action interface | Access and license evidence | Evaluation boundary; feasible reproduction |
|---|---|---|---|---|
| Gemini Robotics / ER | Announcement: 2025-03-12 | Action model versus embodied reasoning model | Selected access; not a verified public weight release | Historical comparison; distinguish perception/reasoning evaluation from robot execution |
| Gemini Robotics On-Device | Announcement: 2025-06-24 | Local action inference | SDK through trusted testing | On-device design study; no unrestricted public checkpoint reproduction established |
| Gemini Robotics 1.5 / ER 1.5 | Announcement: 2025-09-25 | Hierarchical reasoning/action pairing | ER API versus partner VLA access | ER tasks and system interfaces can be studied separately from proprietary control claims |
| Gemini Robotics ER 1.6 | Announcement: 2026-04-14 | Spatial and multi-view embodied reasoning | Service access under provider terms; no public weights verified | Instrument reading or pointing does not measure motor-control success |
| Gemini Robotics 2 | Announcement: 2026-07-30 | Whole-body and dexterous action-model family | VLA available to early-access partners; public downloadable action weights not verified | Analyze reported embodiment/task setup; cannot reproduce full model locally without access |
| Gemini Robotics ER 2 | Model card: published 2026-07-30 | Embodied VLM with text outputs and tool orchestration | API/AI Studio channels; provider terms, not open weights | Reproduce accessible reasoning/interface evaluations with exact service identifier and request records |
| Gemini Robotics On-Device 2 | Model card: published 2026-07-30 | Local VLA, numerical robot-action outputs | Trusted-tester access; no public weight release verified | Local-control claims require authorized artifacts and supported hardware; a stub is only an interface test |
Accessible baselines and 2026 research¶
| Artifact | Dated event and primary evidence | Architecture/action interface | Access and license evidence | Evaluation boundary; feasible reproduction |
|---|---|---|---|---|
| OpenVLA-OFT | Paper: 2025-02-27; v2 April 28 | Parallel continuous action chunks; L1 adaptation recipe | Code, released task checkpoints; MIT code, base-model/checkpoint terms separately | LIBERO/ALOHA paths; preserve camera count, action horizon, normalizer, and checkpoint |
| SmolVLA | Release post: 2025-06-03; paper June 2 | Compact flow action expert | Model card: Apache-2.0; LeRobot code and recipes | Consumer-scale adaptation/replay; CPU execution availability does not imply control-rate feasibility |
| RDT-1B / RDT-170M | Paper: 2024-10-10; repository dates small-model release 2024-10-23 | Diffusion transformer; unified physical action representation | Code, checkpoints and released data; repository states MIT | Bimanual and supported simulation examples; retain external encoder and dataset provenance |
| MolmoAct 2 | Announcement: 2026-05-05; code/data update May 28; repository dates expanded training code June 13 | Flow action expert; optional depth reasoning | Public code, models, data, rollouts; repository states Apache-2.0 | Select a matching adapted policy and pinned evaluator; base, pretraining, and deployment checkpoints differ |
| LingBot-VLA 1.0 | Repository news: code/weights 2026-01-27 | Cross-embodiment action model | Public artifacts; inspect version-specific terms | Useful historical baseline; do not merge its data/schema assumptions with v2 |
| LingBot-VLA 2.0 | Paper: 2026-07-07; repository: July 8 release news, July 25 RoboTwin weights | Expanded body action space; predictive representation learning; MoE action expert | 6B card and code; Apache-2.0 stated | Use public post-training/evaluation artifacts; private pretraining data and GM-100 setup limit full reproduction |
| Qwen-VLA | Paper: 2026-05-28; v2 June 1 | DiT actions/trajectories with embodiment-conditioned prompts | Paper verified; code/weight release and applicable model license not established in this review | Architecture and protocol analysis; no fabricated executable recipe |
| VLAct | Paper: 2026-08-27; v2 2026-10-06 | Transferable action-aware VLM backbone with multiple-head supervision | Code and model zoo; MIT code, inspect individual weight/base-model terms | Adapt backbone or use downstream checkpoint; released heads/settings can differ from headline paper tables |
| DreamZero | Paper: 2026-02-17 | Joint future-video/action generation | Repository: Apache-2.0 code and linked checkpoints; checkpoint/backbone terms separately | Distributed inference, supported simulation and adaptation; full paper runtime requires matching hardware/configuration |
| Continuous Reasoning for VLA | Paper: 2026-05-29; v2 June 8 | Control-aligned reasoning representation | Paper verified; public executable artifacts not established | Study formulation and ablations; course fixture is conceptual, not replication |
| In-Context VLA | Paper: 2026-08-06 | Grounded context plus tool-mediated information acquisition | Paper verified; public executable artifacts not established | Compare method assumptions and reported protocols; use independent fixture for orchestration |
| Figure Helix 02 / Helix 2.5 | Helix 02: 2026-01-27; Helix 2.5: 2026-09-17 | Hierarchical whole-body control; later human-data pretraining study | Official technical announcements; public code/weights not verified | Helix 2.5 reports held-out homes/objects with behavior adaptation elsewhere; unavailable full-system reproduction |
| VLAQuantBench | Paper: 2026-09-21 | Quantization evaluation, not another generalist policy | Paper links public code/configuration/episode records; this review did not execute them | Distinguish calibration, tensor scope, simulated quantization, real kernels, and closed-loop evaluation |
| RoboCasa365 leaderboard | Official table: updated 2026-10-13 | Benchmark snapshot, not a model family | Individual entries have different access; inspect linked artifacts | Compare only within the documented protocol; a ranking does not establish universal superiority |
Unresolved provenance and access notes¶
- GR00T N1.7: the inspected repository describes Apache-2.0 general availability, while the inspected Hugging Face card says EA and links NVIDIA Open Model License. Do not silently apply the repository's license statement to every weight revision. Capture the exact packaged terms and resolve the inconsistency for the artifact you use.
- Physical Intelligence: the publisher-hosted π0.7, π0.6, and memory PDFs returned HTTP 403, but their arXiv records were subsequently verified; full arXiv HTML was read for π0.7 and π0.6. Publisher-download failure does not mean the papers are unavailable. Corresponding public policy weights were not verified.
- VLAct: checkpoint configuration and paper configuration can differ. Reproduce the selected artifact's recorded settings; cite the paper only for the experiments it actually reports.
- Mutable sources: model cards and repositories may change after this snapshot. Their historical event dates and observed access are recorded here, but immutable copies of every page were not archived. Resolve exact revisions before treating this as an executable bill of materials.
- Unverified access is not proof of nonexistence. It means this review did not establish that artifact. An available abstract is sufficient for a dated research pointer, not for asserting weights, code quality, or reproduction success.
How to extend the register¶
Add publication evidence before adding capability claims. Use a primary paper, author repository, model card, or official release. Keep the original event date separate from the date checked. Record the exact evaluation setting and one concrete runnable path, or write that execution is unavailable. Exclude evidence after the snapshot from this historical version; create a later dated review for new releases.
Use the source index for reading order and Lecture 30 for a complete comparison exercise.