motion-language-alignment/motion-language-encoder

Sentence encoder for the anonymous ICRA 2027 submission "Aligning Vehicle Motion with Language"

text_tower_fp16.onnx (669 MB): BGE-large-en-v1.5 backbone fine-tuned jointly with the motion encoder, token mean pooling, L2 normalisation, 1024 to 256 projection, L2 normalisation. Inputs input_ids, attention_mask, token_type_ids (int64, [1, L]); output embedding (float16, [1, 256]). Used by the browser demo at https://motion-language-alignment.github.io/demo/ .