Introduction
An open VLA release can expose inference code and checkpoints while leaving much of the pretraining data, robot-specific action mapping or hardware setup outside the repository. This guide compares OpenVLA, SmolVLA, Physical Intelligence openpi and Octo by license, released artifacts, action representation, compute needs, supported embodiments and physical evaluation rather than by one incompatible benchmark score.
Key findings
- OpenVLA publishes code, checkpoints and fine-tuning workflows for robot-action research, with MIT-licensed software.
- Physical Intelligence openpi publishes multiple policy families, base checkpoints and adaptation examples but explicitly warns that transfer to other robots is not guaranteed.
- The action representation, control frequency and robot configuration matter as much as model architecture when adapting a VLA.
- Inference and fine-tuning compute requirements vary substantially and should be reported with the model version.
- Benchmarks from different robots, datasets and reset rules are not directly comparable; matched physical trials are more useful for deployment decisions.
Open VLA Models: Code, Weights and Adaptation Limits - evidence comparison
The table records what each source establishes and keeps missing data visible.
| System or method | What the evidence establishes | Evidence class | Main unresolved point |
|---|---|---|---|
| OpenVLA | Open code and weights built for robot action generation, with published fine-tuning workflows. | Open VLA project | Benchmarks are not directly comparable. |
| SmolVLA | Hugging Face model and implementation aimed at smaller-scale open robot learning. | Open model ecosystem | Open weights do not expose all pretraining data. |
| OpenPI π models | Physical Intelligence releases code and checkpoints for selected policies, with model-specific terms and hardware demands. | Open research release | Real-robot evidence varies strongly by model and target hardware. |
| Octo | Open generalist policy with multi-robot training and adaptation evidence. | Open policy project | Benchmarks are not directly comparable. |
Evidence classification
- Officially documented specifications, standards or project status stated by the responsible organization.
- Real-system evidence demonstrations or deployments performed on physical hardware under described conditions.
- Company claim a numerical or operational statement reported by the company and not independently audited.
- Simulation or research evidence useful for mechanisms, but not proof of field deployment.
- Insufficient public evidence control mode, trial count, version or operating conditions are missing.
Definition and openness test
An open-source vision-language-action model accepts visual observations and language instructions and outputs robot actions, while publishing reusable code and usually weights. Training data, action tokenization and supported robots may still be partially closed.
How the stack is assembled
Verify repository, model card and license. Record code, weights, data and training recipe separately. Identify action format and control frequency. Check supported embodiments and fine-tuning path. Compare real-robot results only on compatible tasks. Latency, calibration and safety limits can change the result even when the high-level model remains the same.
How to compare open releases
The review method for open source VLA model follows the hardware, software and deployment evidence published by OpenVLA project, Hugging Face, Physical Intelligence. It checks released code, weights, datasets, hardware files, license scope, installation steps and reproducible hardware results and refuses to infer fleet scale, autonomy or reliability from a single edited demonstration.
Reproduction failure modes
A VLA can fail when action normalization or joint ordering does not match the target robot, when inference latency exceeds the task's control budget, or when camera geometry differs from the training setup.
Open code does not remove dataset bias or licensing complexity. Code, checkpoints and training data may use different terms, and benchmark results can depend on fixed cameras, familiar objects or task-specific post-training.
Practical developer uses
Credible applications include Research comparison and fine-tuning, Low-cost manipulation experiments, Cross-embodiment adaptation and Reproducible VLA evaluation. These applications should be described with the robot, task boundary, operator role and environmental constraints. Experimental capability, commercial availability and routine deployment are reported as separate statuses.
What to verify before adoption
Verify the checkpoint license, required model and tokenizer versions, action representation, control frequency, supported camera inputs and the amount of target-robot fine-tuning used in the published result.
Before physical rollout, run repeated trials on the intended robot and report inference latency, task success, interventions, recovery and any robot-specific wrapper or low-level controller that sits outside the VLA.
Limitations and missing information
- Benchmarks are not directly comparable.
- Open weights do not expose all pretraining data.
- Real-robot evidence varies strongly by model and target hardware.
- Specifications, prices, repositories and deployment status can change after publication.
- Benchmarks from different robots or environments are not directly comparable.
Conclusion
Open-source VLA models differ in much more than model size. OpenVLA, SmolVLA, openpi and Octo use different action representations, robot data, supported embodiments and compute budgets. The useful comparison is whether the released checkpoint can be reproduced, adapted to the target action space and evaluated on repeated physical trials with intervention and recovery reported.
Frequently asked questions
How should an open-source VLA model be evaluated?
Record the code and checkpoint license, model version, observation inputs, action representation, control frequency, pretraining-data disclosure, supported robots, inference hardware and adaptation budget. Then compare repeated physical trials on a matched task rather than benchmark percentages from different robots.
What does OpenVLA's openness establish?
OpenVLA publishes software under an MIT license and provides model artifacts and fine-tuning workflows for robot-action research. That makes inspection and adaptation possible, but it does not guarantee that the checkpoint transfers to a new robot without action mapping, calibration and new physical evaluation.
Sources and methodology
Sources for open-source VLA models were rechecked on September 26, 2026 using current OpenVLA, Hugging Face SmolVLA, Physical Intelligence openpi, Octo and LeRobot materials. Licenses, checkpoints, action formats, compute requirements and real-robot evidence are compared separately.
Related TechniaHQRobot guides
Source review notes
Review date recorded July 11, 2026
Supported by the cited sources
- OpenVLA: Open code and weights built for robot action generation, with published fine-tuning workflows.
- SmolVLA: Hugging Face model and implementation aimed at smaller-scale open robot learning.
Not confirmed or incomplete
- Benchmarks are not directly comparable.
- Open weights do not expose all pretraining data.
- Real-robot evidence varies strongly by model and target hardware.
Likely to change quickly
- Commercial availability, prices, model versions and software access.
- Deployment counts, company partnerships and repository maintenance status.
Share this article
Share the current TechniaHQRobot article page.
Continue reading
Open the latest robotics reporting, Physical AI analysis and hardware notes.