| --- |
| library_name: transformers |
| pipeline_tag: robotics |
| base_model: |
| - openbmb/MiniCPM-RobotTrack |
| --- |
| |
| This is a randomly initialized tiny model based on [openbmb/MiniCPM-RobotTrack](https://huggingface.co/openbmb/MiniCPM-RobotTrack) for testing and debugging. |
|
|
| | Feature | Source | This repo | |
| |---|---|---| |
| | Checkpoint tensor elements | 465.65M BF16, 1.13M F32 | 2.39M BF16, 75.13K F32 | |
| | Safetensors weight files | 931.43 MB (model.safetensors) | 5.09 MB (model.safetensors) | |
| | Backbone layers / width / heads | 24 / 1024 / 16 | 4 / 32 / 1 | |
| | Backbone MLP / KV heads | 4096 / 2 | 64 / 1 | |
| | Trajectory funnel linear stages | 6 | 6 | |
| | External visual feature dimension | 1536 | Same | |
| | History / coarse tokens / current fine tokens | 31 / 4 per frame / 64 | Same | |
| | Waypoints / action dimensions | 8 / [x, y, yaw] | Same | |
| | Text vocabulary rows | 73448 | Same | |
| | Stored backbone precision | BF16 | Same | |
| | Input / output | Instruction, DINOv3+SigLIP tokens → trajectory | Same | |
|
|
| RobotTrack accepts fused DINOv3 and SigLIP features, not raw pixels. The source checkpoint does not include those visual encoders; the example validates both required feature streams directly. |
|
|
| | File path | Size | |
| |------|------| |
| | model.safetensors | 5.1MB | |
|
|
| ### Example usage |
|
|
| #### Transformers (✅ verified) |
|
|
| ```python |
| def run_transformers(model_id: str): |
| import torch |
| from transformers import AutoModel, AutoTokenizer |
| |
| model = AutoModel.from_pretrained(model_id, trust_remote_code=True, dtype=torch.bfloat16).eval().cuda() |
| tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) |
| if tokenizer.pad_token_id is None: |
| tokenizer.pad_token = tokenizer.eos_token |
| text = tokenizer("Follow the person in the blue shirt.", return_tensors="pt") |
| coarse = torch.zeros(1, 31 * 4, 1536, device=model.device) |
| coarse_time = torch.arange(31, device=model.device).repeat_interleave(4).unsqueeze(0) |
| fine = torch.zeros(1, 64, 1536, device=model.device) |
| fine_time = torch.full((1, 64), 31, dtype=torch.long, device=model.device) |
| output = model( |
| input_ids=text.input_ids.to(model.device), |
| attention_mask=text.attention_mask.to(model.device), |
| coarse_tokens=coarse, |
| coarse_time_indices=coarse_time, |
| fine_tokens=fine, |
| fine_time_indices=fine_time, |
| ) |
| assert output.trajectories.shape == (1, 8, 3) |
| assert torch.isfinite(output.trajectories).all() |
| print(output.trajectories.shape) |
| |
| |
| if __name__ == "__main__": |
| run_transformers(model_id="tiny-random/minicpm-robot-track") |
| ``` |
|
|
| ### Creating this repo |
|
|
| - Code: [.meta/create_model.py](.meta/create_model.py) |
| - Model structure: [.meta/model.log](.meta/model.log) |
| - Environment: [.meta/env.json](.meta/env.json) |
|
|
| ### Change log |
|
|
| - 2026-10-27: Created a tiny MiniCPM RobotTrack visual-token trajectory policy. |
|
|
| *This repo is generated by GPT-6.1-Sol.* |
