Peking University · robot AI model
LDA-1B
China · 2026
Runs on G1 by Galbot · G1 by Unitree RoboticsSource and note: G1 (Galbot), G1 (Unitree Robotics)B pku-epic.github.io/LDA/ · Retrieved 13 November 2026
Sources retrieved: 13 November 2026 (A, B, C, D)
- What it controls
- Arms and handsSource and note: What it controlsA github.com/jiangranlv/latent-dynamics-action · Retrieved 13 November 2026Note. Code repository: “Predicts denoised 16-step action chunks”. Project page, robot platforms: “Galbot G1 equipped with a standard two-finger parallel gripper for basic grasping tasks”; “Galbot G1 fitted with the SharpaWave dexterous hand (22 DoF) for fine manipulation”; “Unitree G1 mounted with the BrainCo dexterous hand (10 DoF) and a Zed Mini camera”.
- Inputs
- Camera LanguageSource and note: InputsA github.com/jiangranlv/latent-dynamics-action · Retrieved 13 November 2026Note. Code repository: the language and vision encoder is “Qwen3-VL (4B) → extracts semantics information”; image tokens are DINOv3 patch embeddings.
- Access
- Open weights (download)Source and note: AccessD huggingface.co/api/models/Wayer2/LDA-pretrain · Retrieved 13 November 2026Note. Hugging Face repository metadata list the file “LDA-pretrain.pt” and mark the repository as not gated.
- Licence type
- Free, commercial useSource and note: Licence typeD huggingface.co/api/models/Wayer2/LDA-pretrain · Retrieved 13 November 2026Note. Pretrained checkpoint under “license:apache-2.0”, not gated. The project page lists the training data as coming soon.
Specifications
8 of 14 stated 6 not stated
3 fields hidden 3 not-stated fields with a note stay ·
How to read the figuresShow guideClose
- not statedNo source
- yesStated by a source
- 0Reported as zero
- ASource
Select a source marker for source details and notes.
Release1/1 disclosed
Released12 February 2026
Source and note: ReleasedC arxiv.org/abs/2602.12215 · Retrieved 13 November 2026Note. arXiv listing: “Submitted on 12 Feb 2026”.
Input and output2/4 disclosed
Model typeNot stated
Source and note: Model typeB pku-epic.github.io/LDA/ · Retrieved 13 November 2026Note. The paper title calls it a Latent Dynamics Action Model. Project page: “LDA-1B is a dynamics-centric robot foundation model trained on 30k+ hours of heterogeneous embodied data”. Code repository: “Single MMDiT backbone jointly predicts future visual features (`DINOv3` tokens) and 16-step action chunks”.
InputsCamera Language
Source and note: InputsA github.com/jiangranlv/latent-dynamics-action · Retrieved 13 November 2026Note. Code repository: the language and vision encoder is “Qwen3-VL (4B) → extracts semantics information”; image tokens are DINOv3 patch embeddings.
What it controlsArms and hands
Source and note: What it controlsA github.com/jiangranlv/latent-dynamics-action · Retrieved 13 November 2026Note. Code repository: “Predicts denoised 16-step action chunks”. Project page, robot platforms: “Galbot G1 equipped with a standard two-finger parallel gripper for basic grasping tasks”; “Galbot G1 fitted with the SharpaWave dexterous hand (22 DoF) for fine manipulation”; “Unitree G1 mounted with the BrainCo dexterous hand (10 DoF) and a Zed Mini camera”.
Degrees of freedom controlledNot stated
Size and speed0/2 disclosed
ParametersNot stated
Source and note: ParametersC arxiv.org/abs/2602.12215 · Retrieved 13 November 2026Note. The developer gives parts, not a total. Paper abstract: LDA-1B employs a multi-modal diffusion transformer, “enabling stable training at the 1B-parameter scale”. Code repository: the language and vision encoder is “Qwen3-VL (4B) → extracts semantics information”.
Control rateNot stated
Where it runs and what it is built on0/2 disclosed
On robot or cloudNot stated
Base modelNot stated
Source and note: Base modelA github.com/jiangranlv/latent-dynamics-action · Retrieved 13 November 2026Note. Code repository: the architecture uses “Qwen3-VL (4B) → extracts semantics information” with a frozen DINOv3 vision encoder; the developer does not name a base model.
Openness3/3 disclosed
AccessOpen weights (download)
Source and note: AccessD huggingface.co/api/models/Wayer2/LDA-pretrain · Retrieved 13 November 2026Note. Hugging Face repository metadata list the file “LDA-pretrain.pt” and mark the repository as not gated.
Licence documentApache-2.0
Source and note: Licence documentD huggingface.co/api/models/Wayer2/LDA-pretrain · Retrieved 13 November 2026Note. Hugging Face repository metadata for the pretrained checkpoint: “license:apache-2.0”. The code repository README carries an MIT licence badge, which covers the code.
Licence typeFree, commercial use
Source and note: Licence typeD huggingface.co/api/models/Wayer2/LDA-pretrain · Retrieved 13 November 2026Note. Pretrained checkpoint under “license:apache-2.0”, not gated. The project page lists the training data as coming soon.
The developer’s own claims2/2 disclosed
Learns new tasksNeeds retraining
Source and note: Learns new tasksC arxiv.org/abs/2602.12215 · Retrieved 13 November 2026Note. Paper abstract: “LDA-1B enables data-efficient fine-tuning”.
Works across robotsYes
Source and note: Works across robotsA github.com/jiangranlv/latent-dynamics-action · Retrieved 13 November 2026Note. Code repository: “Pre-trained on multi embodiments (Agibot, Unitree-G1, Human, etc.)”. Project page, robot platforms: “Galbot G1 equipped with a standard two-finger parallel gripper for basic grasping tasks”; “Unitree G1 mounted with the BrainCo dexterous hand (10 DoF) and a Zed Mini camera”.Yes only if the developer itself says the model runs on different robot models, for example robots from two manufacturers or two of one manufacturer’s models. Several units of the same robot do not count.
Notes
The project page credits Peking University and Galbot, together with CASIA, BAAI, Tsinghua University, Sun Yat-sen University and NVIDIA. The paper title calls it a Latent Dynamics Action Model, and the abstract says: “We introduce LDA-1B, a robot foundation model that scales through universal embodied data ingestion by jointly learning dynamics, policy, and visual forecasting”.
Sources
One line per source, not per figure: a source has one retrieval date.
- Agithub.com/jiangranlv/latent-dynamics-actionRetrieved 13 November 2026
- Bpku-epic.github.io/LDA/Main source. Retrieved 13 November 2026
- Carxiv.org/abs/2602.12215Retrieved 13 November 2026
- Dhuggingface.co/api/models/Wayer2/LDA-pretrainRetrieved 13 November 2026