Source-linked model matrix
VLA model comparison
Compare vision-language-action models by inputs, code access, training data and source-specific results. Start with contact feedback when your task depends on touch.
By RoboSkin.ai Editorial Team · Source review:
Methodology and scope
Inclusion: published robot policies that connect visual observations and instructions to action, plus explicitly labeled related planners and policy methods needed to understand the field. PaLM-E is a related planner; original Diffusion Policy is a related visuomotor policy. Inclusion does not imply that every row is an end-to-end VLA.
Sources: arXiv papers, official project pages and author-linked repositories. Each factual cell links its evidence. Year means first arXiv submission; access describes the source-review date. Updates are manual and event-driven: records are revised when a release or correction is checked against primary sources. There is no automatic refresh.
“Open code” describes a released implementation, not unrestricted weights, data or a fully reproducible deployment. Partial releases and unknown availability have separate filters. “not found” records an unresolved field, not a zero value. A tactile checkmark includes documented contact-force sensing and identifies variant-only or modular integration in the row.
Benchmark scores are examples from different tasks, robots, datasets and metrics, not a shared leaderboard. Parameter counts also differ in whether backbones or whole systems are counted. Follow the source before comparing numbers, and use the existing tactile integration evidence map for deeper mechanism-level context.
Compare models and contact inputs
18 of 18 models
Scroll horizontally to read every column. Sort by model, organization, year or parameter count using the column headings. Parameter sorting uses the largest sortable size stated in each row and places unknown totals last; each row retains the source’s counting scope. This is not a compute comparison. Benchmark cells retain their own protocols.
| Paper link | Code link | Input modalities | Open-source status | Benchmark scores | Training dataset | One-line description | ||||
|---|---|---|---|---|---|---|---|---|---|---|
| VelaVLADetailed evidence | Fudan University, Shanghai Innovation Institute, CUHK and NovaXBotSource ↗ | Paper ↗ for Vela | not found | not found — complete model totalSource ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Vision, language and robot state; no live tactile input documented in the reviewed architecture.Source ↗ | Closed / unreleasedUnreleased: the official project labels code “Coming soon”.Source ↗ | 45.3% LIBERO-X / 49.7% EBench success; separate simulation benchmarks.Source ↗ | Public and private cross-embodiment robot trajectories, followed by task-specific post-training.Source ↗ | Represents future actions with adaptive-horizon curves instead of fixed-rate action chunks.Source ↗ | |
| ViTaRVLADetailed evidence | not found — institutional affiliation in the reviewed paper / projectSource ↗ | Paper ↗ for ViTaR | not found | not found — full base-plus-adapter totalSource ↗ | Yes: VisionYes: LanguageYes: Tactile / force Real-time visuo-tactile observations select and scale residual corrections around frozen OpenVLA-OFT actions.Source ↗ | Closed / unreleasedUnreleased: the official project labels code “Coming soon”.Source ↗ | 61.3% UniVTAC average; 48.3% physical-robot average. Separate simulation and physical cohorts.Source ↗ | UniVTAC task demonstrations and local decision / branch data for residual adaptation.Source ↗ | Constrains touch to bounded residual selection and scaling around a frozen foundation policy.Source ↗ | |
| ForceVLA2VLA | Shanghai AI Laboratory, Tongji University, Shanghai Jiao Tong University and collaborators listed in the paperSource ↗ | Paper ↗ for ForceVLA2 | not found | not found — full adapted model totalSource ↗ | Yes: VisionYes: LanguageYes: Tactile / force Images, prompts, proprioception and real-time interaction force. Force feedback is not a surface tactile image.Source ↗ | Closed / unreleasedUnreleased: the official project labels code and dataset “coming soon”.Source ↗ | 66.0% mean success over 5 physical contact-rich tasks, 20 trials per task; paper Table 1.Source ↗ | ForceVLA2-Dataset: 1,000 trajectories with images, prompts, state and force.Source ↗ | Combines force-aware prompting and hybrid force–position action control for contact-rich tasks.Source ↗ | |
| OmniVTLAVLA | Paxini Tech. and Shanghai Jiao Tong UniversitySource ↗ | Paper ↗ for OmniVTLA | not found | not found — complete model total in the reviewed sourceSource ↗ | Yes: VisionYes: LanguageYes: Tactile / force Vision, language and tactile signals, with separate paths for visual and force-based tactile sensors.Source ↗ | not foundnot found — an official implementation / checkpoint link in the reviewed paper.Source ↗ | 96.9% gripper / 100% dexterous-hand pick-and-place success in paper v1; different configurations, not one aggregate.Source ↗ | ObjTac for semantic tactile alignment, existing visuo-tactile data and robot task demonstrations.Source ↗ | Aligns heterogeneous tactile signals with vision and language for contact-aware action generation.Source ↗ | |
| VLA-TouchVLADetailed evidence | National University of SingaporeSource ↗ | Paper ↗ for VLA-Touch | Code ↗ for VLA-Touch | not found — full modular system total; do not substitute the RDT backbone sizeSource ↗ | Yes: VisionYes: LanguageYes: Tactile / force GelSight touch informs language-level planning and action refinement around a non-tactile RDT base policy.Source ↗ | Partial releasePartial open code (MIT), controller/data links; the modified RDT inference script is still described as a future release.Source ↗ | 12/20 Wipe completions for the tactile interpolant controller; manipulation component evaluation, not full planning-plus-execution success.Source ↗ | Task-specific tactile demonstrations and pretrained RDT / Octopi components; no single unified training corpus.Source ↗ | Adds tactile feedback at planning and execution levels without fine-tuning the base VLA.Source ↗ | |
| SmolVLAVLA | Hugging Face with Sorbonne University, valeo.ai and ENS Paris-Saclay collaboratorsSource ↗ | Paper ↗ for SmolVLA | Code ↗ for SmolVLA | 450M (main model in paper v1)Source ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Camera images, language and robot state; no tactile input in the reported interface.Source ↗ | Open codeOpen LeRobot code (Apache-2.0), with model and dataset releases documented by the authors.Source ↗ | 87.3% LIBERO average for the 0.45B model in paper v1, Table 2; simulation, without robot-data pretraining in that experiment.Source ↗ | Community-collected LeRobot demonstrations; simulation experiments train on their benchmark datasets separately.Source ↗ | Explores a compact VLA and asynchronous inference using accessible robot hardware and community data.Source ↗ | |
| VTLAVLA | Samsung R&D Institute China–Beijing, BAAI and Institute of Automation, Chinese Academy of SciencesSource ↗ | Paper ↗ for VTLA | not found | Qwen2-VL 7B backbone; complete model total not foundSource ↗ | Yes: VisionYes: LanguageYes: Tactile / force Visual, tactile and language inputs; temporally encoded tactile observations for insertion.Source ↗ | not foundnot found — public implementation or checkpoint in the reviewed paper and project.Source ↗ | 95% triangle / 100% pentagon physical insertion at 0.6 mm clearance, 20 trials per shape; paper Table 7.Source ↗ | Authors’ simulated vision–tactile–action–instruction insertion dataset and preference data.Source ↗ | Applies preference learning to a vision–tactile–language policy for contact-rich insertion.Source ↗ | |
| π0.5 (pi0.5)VLA | Physical IntelligenceSource ↗ | Paper ↗ for π0.5 (pi0.5) | Code ↗ for π0.5 (pi0.5) | not found — exact total for the reviewed releaseSource ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Images, language and robot state; the reference policy does not document live tactile input.Source ↗ | Open codeOpen code (Apache-2.0) and base checkpoints; the repository documents the supported flow-matching head.Source ↗ | not found — a single comparable aggregate for the varied home-task evaluation; inspect per-task results.Source ↗ | Co-training across robot embodiments, high-level semantic prediction and web-derived data.Source ↗ | Studies open-world generalization through heterogeneous co-training and hierarchical action prediction.Source ↗ | |
| GR00T N1VLADetailed evidence | NVIDIASource ↗ | Paper ↗ for GR00T N1 | Code ↗ for GR00T N1 | 2.2B (GR00T-N1-2B release)Source ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Vision, language and robot state; no tactile input documented for the N1 release.Source ↗ | Open codeOpen code (Apache-2.0) and model links; consult the N1 checkpoint’s separate terms and version.Source ↗ | 32.1% RoboCasa success with 100 demonstrations per task, paper Table 2; simulation only.Source ↗ | Real-robot trajectories, human video and synthetic data.Source ↗ | Couples a vision-language reasoning module to a diffusion action module for humanoid control.Source ↗ | |
| OpenVLA-OFTVLA | Stanford UniversitySource ↗ | Paper ↗ for OpenVLA-OFT | Code ↗ for OpenVLA-OFT | 7B OpenVLA base; exact adapted total not foundSource ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Images, instructions and optional proprioception; no tactile input in the reference recipe.Source ↗ | Open codeOpen code (MIT) and checkpoints; inherited OpenVLA base-model restrictions remain relevant.Source ↗ | 97.1% mean success across 4 LIBERO simulation suites in the OFT paper.Source ↗ | OpenVLA initialization; LIBERO and ALOHA task demonstrations for separate fine-tuning experiments.Source ↗ | Optimizes VLA fine-tuning with parallel decoding, action chunking and continuous-action regression.Source ↗ | |
| π0 (pi0)VLADetailed evidence | Physical IntelligenceSource ↗ | Paper ↗ for π0 (pi0) | Code ↗ for π0 (pi0) | 3.3B total in the original paperSource ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Images, language and proprioception; no tactile input in the original policy.Source ↗ | Open codeOpen code (Apache-2.0) and base checkpoints through openpi; upstream model terms also apply.Source ↗ | not found — a single original-paper aggregate across its distinct zero-shot and fine-tuned task protocols.Source ↗ | Cross-embodiment robot demonstrations, including single-arm, dual-arm and mobile platforms.Source ↗ | Combines pretrained vision-language knowledge with flow matching for continuous action chunks.Source ↗ | |
| RDT (RDT-1B)VLA | Tsinghua UniversitySource ↗ | Paper ↗ for RDT (RDT-1B) | Code ↗ for RDT (RDT-1B) | 1.2B (diffusion model)Source ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Images, language and robot state; the original model does not consume tactile feedback.Source ↗ | Open codeOpen code, model weights and released data under MIT according to the official repository.Source ↗ | 50% Wash Cup success with unseen cup 2 in the paper’s ablation table; one physical-task condition, not an overall score.Source ↗ | Multi-robot pretraining collection and a separately collected bimanual fine-tuning dataset.Source ↗ | Scales a diffusion transformer and unified action representation for bimanual manipulation.Source ↗ | |
| OpenVLAVLADetailed evidence | Stanford, UC Berkeley, Toyota Research Institute, Google DeepMind, Physical Intelligence and MITSource ↗ | Paper ↗ for OpenVLA | Code ↗ for OpenVLA | 7BSource ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force RGB observations and instructions mapped to discretized actions; no live tactile input in the base release.Source ↗ | Open codeOpen code (MIT) and weights; pretrained checkpoints inherit base-model restrictions.Source ↗ | 76.5% LIBERO mean success for the OpenVLA reference in the OFT paper; a later benchmark evaluation, not the original real-robot aggregate.Source ↗ | 970k real-world demonstrations from an Open X-Embodiment mixture.Source ↗ | Makes a pretrained vision-language backbone available for robot action prediction and adaptation.Source ↗ | |
| OctoVLADetailed evidence | UC Berkeley, Stanford, Carnegie Mellon and Google DeepMindSource ↗ | Paper ↗ for Octo | Code ↗ for Octo | 27M / 93M (Small / Base)Source ↗ | Yes: VisionYes: LanguageYes: Tactile / force Camera observations, language or goal images. Contact input only in the force/torque fine-tuned peg-insertion variant; base checkpoints do not consume touch.Source ↗ | Open codeOpen code (MIT) with base checkpoints and fine-tuning tools.Source ↗ | 0.72 average success in the project’s fine-tuning evaluation; separate from zero-shot results.Source ↗ | 800k trajectories from Open X-Embodiment, plus target-task data for adaptation.Source ↗ | Provides an adaptable generalist policy with a diffusion head and configurable observation / action spaces.Source ↗ | |
| RT-2VLADetailed evidence | Google DeepMindSource ↗ | Paper ↗ for RT-2 | not found | 12B / 55B (project’s main evaluated variants)Source ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Robot images and natural-language instructions; no tactile input in the described interface.Source ↗ | Closed / unreleasedClosed research model; public policy implementation and weights not found. Language Table is a dataset release, not RT-2 code.Source ↗ | not found — a single aggregate shared by the variants and evaluation categories; use the paper’s per-condition results.Source ↗ | Robot demonstrations co-fine-tuned with web-scale vision-language tasks.Source ↗ | Represents robot actions as text tokens to transfer vision-language knowledge into control.Source ↗ | |
| Diffusion PolicyRelated policy | Columbia University, Toyota Research Institute and MITSource ↗ | Paper ↗ for Diffusion Policy | Code ↗ for Diffusion Policy | not found — no single size for the method’s architecture variantsSource ↗ | Yes: VisionNot documented: LanguageNot documented: Tactile / force Visual / state-conditioned action diffusion; language and tactile input are not part of the original evaluated interface.Source ↗ | Open codeOpen code (MIT), training configurations, data and checkpoint links.Source ↗ | 46.9% average improvement over prior methods across 12 tasks / 4 benchmarks, as reported in the abstract; this is an improvement, not a success rate.Source ↗ | Benchmark demonstrations and authors’ real-world demonstrations; task-specific training.Source ↗ | Generates action sequences by conditional denoising; an influential visuomotor policy rather than a language-conditioned VLA.Source ↗ | |
| PaLM-ERelated plannerDetailed evidence | Google Research and TU BerlinSource ↗ | Paper ↗ for PaLM-E | not found | 562B (largest reported variant)Source ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Images, text and continuous state encodings; outputs text / plans for a separate low-level policy.Source ↗ | Closed / unreleasedClosed research model; public model code and weights not found in the reviewed project.Source ↗ | not found — a single robot-action aggregate applicable to all variants; the paper separates embodied and question-answering evaluations.Source ↗ | Embodied-task data mixed with language, vision and vision-language data.Source ↗ | Grounds a language model in continuous observations for embodied reasoning and high-level planning.Source ↗ | |
| RT-1VLA | Google Research / Robotics at GoogleSource ↗ | Paper ↗ for RT-1 | Code ↗ for RT-1 | 35MSource ↗ | Yes: VisionYes: LanguageNot documented: Tactile / force Camera history and language instruction; no tactile input in the described model.Source ↗ | Open codeOpen code (Apache-2.0). Code access does not establish release of the full original training corpus or all checkpoints.Source ↗ | 97% success on the authors’ seen-task evaluation; not an unseen-task score.Source ↗ | Authors’ real-robot demonstration collection.Source ↗ | Scales instruction-conditioned robotic control with a transformer that emits discretized actions.Source ↗ |
Evidence and corrections
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.