Shanghai AI Laboratory is an artificial intelligence research institution. Its IPEC community releases EO-1 and embodied AI research resources.

Overview
Related models
Publicly sourced records only; these are not an exhaustive list of supported robots or evidence of factory installation. Evidence levels describe public documentation, not model capability.
In-house models · 3
| Model | Type | Current version | Openness | Related robots | Latest release |
|---|---|---|---|---|---|
| EO-1 | TypeAction policy | Current version Awaiting verification | Openness Code and weights | Related robots | Latest release |
| InternVLA | TypeAwaiting verification | Current version Awaiting verification | Openness Code and weights | Related robots None recorded | Latest release Awaiting verification |
| X-VLA | TypeAction policy | Current version Awaiting verification | Openness Code and weights | Related robots | Latest release |
Model participation
Version scope and roles for these associations await verification.
Timeline
- Model released
InternVLA-A1.5 introduces training-only video foresight and an action-only deployment path
The first A1.5 paper on July 6 introduces a native Qwen3.5 VLM with a smaller unified expert and learnable latent-foresight tokens supervised by frozen WAN during training. Its deployment design permits continuous actions without the WAN video branch. In the current implementation, explicitly choose optimized plus action_loss_only=true; saved standard configs do not automatically select that path. The archive date does not establish the first code or weight opening.
- Model released
EO-1 control policy merges into LeRobot
LeRobot PR3403merged on May6,2026, adding standard robot-control training, processors, save/load and evaluation. It explicitly excludes full upstream interleaved pretraining and multimodal-generation workflows. READMElinksPR1971andclaimsmerged, but1971wasclosed; the official merged PR is3403.
- Model released
ST4VLA follow-up analyzes spatially guided training within the M1 framework
The February10 first ST4VLA archive provides a separate 12-author conference paper with additional spatial-training analysis. Its Google VM/VA and WidowX reports differ from the earlier report; mixed changes are recorded separately, while identical LIBERO reports are not counted as independent replications. This does not establish a new released checkpoint generation or replace the original contributor list.
- Model released
InternVLA-A1 releases RoboTwin 2.0 fine-tuned weights
The January 23 entry announces InternVLA-A1-3B-RoboTwin for the 50-task RoboTwin 2.0 benchmark. The pinned card reports Easy 89.40% and Hard 89.64% under the stated simulation setup. These are developer reports, not independent current-revision reproduction or real-robot transfer evidence.
- Model released
InternVLA-A1 adds public pretraining code and instructions
The January 14 entry announces pretraining code and guidelines for InternVLA-A1-3B. The tutorial covers InternData-A1 preparation and per-embodiment delta-action statistics; released base weights and dataset-specific permissions remain separate matters.
- Model released
InternVLA-A1 announces its LeRobot V3.0 codebase
The official News entry dates the InternVLA-A1 codebase for the LeRobot V3.0 ecosystem to January 5. The first arXiv paper submission is also January 5, although README labels that paper entry January 7. This supports a specific implementation/workflow announcement; it does not establish the first overall A1 model, code or weight opening. Current code terms remain noncommercial.
- Model released
DualVLN paper documents an RGB dual-system navigation design
The December9 archived DualVLN paper documents a joint RGB navigation configuration with explicit pixel goals, latent features and a fast trajectory policy. Its experiments separate continuous and physical navigation, including dynamic-obstacle limitations. This is the paper date, not a proved first weight-opening date.
- Model released
InternVLA-M1 technical report documents spatially guided robot training
The October15 first archived report documents M1 spatial-grounding pretraining, spatially prompted action post-training and its task-specific evaluations. This dates the document, not the first overall model, code or weight opening. Root VLM and complete policy bundles are distinct.
- Model released
EO-1 introduces unified interleaved robot reasoning and control
The August28first paper introduces EO-1 and EO-Data1.5M under the original EmbodiedOneVision paper title, combining autoregressive vision/language and continuous flow-matching actions in a unified model. It is the research announcement; a weight/code opening on that exact date is not independently established. Laterpaper title/revisions and today’s pinned artifact remain distinct.
- Model released
EO-1 introduces unified interleaved robot reasoning and control
The August28first paper introduces EO-1 and EO-Data1.5M under the original EmbodiedOneVision paper title, combining autoregressive vision/language and continuous flow-matching actions in a unified model. It is the research announcement; a weight/code opening on that exact date is not independently established. Laterpaper title/revisions and today’s pinned artifact remain distinct.
- Model released
InternVLA-M1 technical report documents spatially guided robot training
The October15 first archived report documents M1 spatial-grounding pretraining, spatially prompted action post-training and its task-specific evaluations. This dates the document, not the first overall model, code or weight opening. Root VLM and complete policy bundles are distinct.
- Model released
DualVLN paper documents an RGB dual-system navigation design
The December9 archived DualVLN paper documents a joint RGB navigation configuration with explicit pixel goals, latent features and a fast trajectory policy. Its experiments separate continuous and physical navigation, including dynamic-obstacle limitations. This is the paper date, not a proved first weight-opening date.
- Model released
InternVLA-A1 announces its LeRobot V3.0 codebase
The official News entry dates the InternVLA-A1 codebase for the LeRobot V3.0 ecosystem to January 5. The first arXiv paper submission is also January 5, although README labels that paper entry January 7. This supports a specific implementation/workflow announcement; it does not establish the first overall A1 model, code or weight opening. Current code terms remain noncommercial.
- Model released
InternVLA-A1 adds public pretraining code and instructions
The January 14 entry announces pretraining code and guidelines for InternVLA-A1-3B. The tutorial covers InternData-A1 preparation and per-embodiment delta-action statistics; released base weights and dataset-specific permissions remain separate matters.
- Model released
InternVLA-A1 releases RoboTwin 2.0 fine-tuned weights
The January 23 entry announces InternVLA-A1-3B-RoboTwin for the 50-task RoboTwin 2.0 benchmark. The pinned card reports Easy 89.40% and Hard 89.64% under the stated simulation setup. These are developer reports, not independent current-revision reproduction or real-robot transfer evidence.
- Model released
ST4VLA follow-up analyzes spatially guided training within the M1 framework
The February10 first ST4VLA archive provides a separate 12-author conference paper with additional spatial-training analysis. Its Google VM/VA and WidowX reports differ from the earlier report; mixed changes are recorded separately, while identical LIBERO reports are not counted as independent replications. This does not establish a new released checkpoint generation or replace the original contributor list.
- Model released
EO-1 control policy merges into LeRobot
LeRobot PR3403merged on May6,2026, adding standard robot-control training, processors, save/load and evaluation. It explicitly excludes full upstream interleaved pretraining and multimodal-generation workflows. READMElinksPR1971andclaimsmerged, but1971wasclosed; the official merged PR is3403.
- Model released
InternVLA-A1.5 introduces training-only video foresight and an action-only deployment path
The first A1.5 paper on July 6 introduces a native Qwen3.5 VLM with a smaller unified expert and learnable latent-foresight tokens supervised by frozen WAN during training. Its deployment design permits continuous actions without the WAN video branch. In the current implementation, explicitly choose optimized plus action_loss_only=true; saved standard configs do not automatically select that path. The archive date does not establish the first code or weight opening.
References
Publisher not provided · Type unconfirmed
Publication date unknown · Accessed: 2026-10-27
Linked facts: 1
Source category does not verify the claims it contains.