RT-1
Google Research / Robotics at Google
Scales instruction-conditioned robotic control with a transformer that emits discretized actions. Source ↗
Part of VLA robotics.
Trace the development of vision-language-action models, related policy methods and tactile extensions, from RT-1 to the most recent paper in this reviewed selection.
By RoboSkin.ai Editorial Team · Source review:
Dates below are the first arXiv submission dates linked in each entry, rather than code-release or conference dates. This curated chronology shares its records with the model comparison matrix. It includes related planners and policies where labeled and is not an exhaustive claim about the newest model worldwide.
Google Research / Robotics at Google
Scales instruction-conditioned robotic control with a transformer that emits discretized actions. Source ↗
Grounds a language model in continuous observations for embodied reasoning and high-level planning. Source ↗
Columbia University, Toyota Research Institute and MIT
Generates action sequences by conditional denoising; an influential visuomotor policy rather than a language-conditioned VLA. Source ↗
Represents robot actions as text tokens to transfer vision-language knowledge into control. Source ↗
UC Berkeley, Stanford, Carnegie Mellon and Google DeepMind
Provides an adaptable generalist policy with a diffusion head and configurable observation / action spaces. Source ↗
Stanford, UC Berkeley, Toyota Research Institute, Google DeepMind, Physical Intelligence and MIT
Makes a pretrained vision-language backbone available for robot action prediction and adaptation. Source ↗
Scales a diffusion transformer and unified action representation for bimanual manipulation. Source ↗
Combines pretrained vision-language knowledge with flow matching for continuous action chunks. Source ↗
Optimizes VLA fine-tuning with parallel decoding, action chunking and continuous-action regression. Source ↗
Couples a vision-language reasoning module to a diffusion action module for humanoid control. Source ↗
Studies open-world generalization through heterogeneous co-training and hierarchical action prediction. Source ↗
Samsung R&D Institute China–Beijing, BAAI and Institute of Automation, Chinese Academy of Sciences
Applies preference learning to a vision–tactile–language policy for contact-rich insertion. Source ↗
Hugging Face with Sorbonne University, valeo.ai and ENS Paris-Saclay collaborators
Explores a compact VLA and asynchronous inference using accessible robot hardware and community data. Source ↗
National University of Singapore
Adds tactile planning and action refinement without tactile fine-tuning of the base VLA; the RDT baseline is still aligned to the robot and fine-tuned per task without touch. Source ↗
Paxini Tech. and Shanghai Jiao Tong University
Aligns heterogeneous tactile signals with vision and language for contact-aware action generation. Source ↗
Combines force-aware prompting and hybrid force–position action control for contact-rich tasks. Source ↗
not found — institutional affiliation in the reviewed paper / project
Constrains touch to bounded residual selection and scaling around a frozen foundation policy. Source ↗
Fudan University, Shanghai Innovation Institute, CUHK and NovaXBot
Represents future actions with adaptive-horizon curves instead of fixed-rate action chunks. Source ↗
Xspark AI, Tsinghua University, HKU, HKUST (Guangzhou) and Peking University
Touch serves as training supervision and optional deployment input. The authors describe one checkpoint using tactile placeholders for vision-only operation or live touch for tactile-conditioned operation; training supervision is not a requirement for a sensor at every deployment. Source ↗
Reading guide reviewed November 4, 2026. Original catalog review dates remain above; no model execution.
This timeline orders first arXiv submissions. It does not order the first public announcement, accepted conference version, code release or checkpoint release. On November 4, 2026, all 19 displayed first-submission dates matched the official arXiv metadata. That check validates the chronology field; it is not a fresh reproduction or complete artifact audit.
Diffusion Policy illustrates the distinction: its first arXiv submission was March 7, 2023, while the current arXiv record identifies a March 14, 2024 extended journal version. The original pi0 entry starts on October 31, 2024; its arXiv history also contains a January 8, 2026 revision. A later revision can update evidence without making the original method a newly released model.
| Event | Evidence to preserve | What it cannot date |
|---|---|---|
| First preprint submission | Original identifier and v1 timestamp | Every earlier talk, demo or private result |
| Paper revision | Version, revised date and changed claims | The first release of an unchanged method |
| Code or weight release | Repository/tag or checkpoint revision and actual files | A complete working robot deployment |
| Conference or journal publication | Publisher/venue record and version relationship | When code became available |
| Directory review | Which fields were checked and unresolved access | A new source event or new experiment |
The entries trace several overlapping questions: conditioning actions on instructions, generating actions with different algorithms, adapting across robots, and incorporating contact feedback. Related planners and visuomotor methods are labelled because they illuminate those interfaces; their inclusion does not turn every entry into an end-to-end VLA.
Use chronology to locate the source version, then use the comparison matrix for configuration and evidence. A newer paper is not automatically more deployable, more accurate or more relevant to a contact task. The selected set is intentionally bounded and should not be described as every model released worldwide.
This is an editorial reading aid based on the linked primary sources. Results are author-reported; RoboSkin has not reproduced these model evaluations. Artifact availability and license terms can change.