UnifoLM-ER-1-4B

Project Page

Across 16 multimodal perception and understanding benchmarks, UnifoLM-ER-1 leads open-source models on seven and delivers overall performance comparable to leading proprietary models. Built on Qwen3-VL-4B, UnifoLM-ER-1 is trained on more than 5 million samples spanning image point prediction, object detection, multi-image reasoning, 2D trajectory prediction, 3D object detection, and multi-image spatial question answering. These data are co-trained with general image-text data, preserving broad vision-language capabilities while substantially improving spatial understanding and reasoning in embodied environments.

Benchmark Results

Model	Open Source	RoboVQA	Ego-Plan2	RefSpatial-Bench	Where2Place	Pixmo-Point	BLINK	CV-Bench	EmbSpatial	RoboSpatial	SAT	VSI-Bench	VSR	ERQA	RealWorldQA	MME	MMMU_VAL
UnifoLM-ER-1-4B	Yes	62.4	55.1	61.7	82.0	73.8	93.4†	88.6	88.9	73.1	76.0	54.2	88.1	50.0	69.8	2223.3	54.7
RoboBrain2.0-7B*	Yes	30.0	33.23	42.2	63.6	54.7	83.9	85.7	76.3	54.2	75.3	36.1	84.0	—	69.4	2057.4	44.4
Robix-7B*	No	63.6	—	—	41.9	29.5	87.6	86.5	77.4	—	71.1	44.6	83.3	42.5	70.7	2332.8	—
Pelican-7B*	Yes	31.8	33.7	22.3	57.3	20.4	—	79.4	73.2	57.5	52.0	52.8	82.2	39.8	69.3	2141.9	51.1
Cosmos-R1-7B*	Yes	38.8	26.0	5.6	2.9	8.2	—	76.7	68.9	42.4	82.7	25.4	82.4	—	67.6	2157.4	37.4
Cosmos3-Super-64B*	Yes	—	—	57.0	71.0	—	90.3†	88.0	—	70.0	—	60.9	—	51.2	—	—	—
Qwen3-VL-4B*	Yes	47.7	40.7	46.6	63.0	48.3	85.0†	85.1	79.6	61.7	68.7	59.3	81.6	41.3	71.0	2325.2	57.8
Qwen3-VL-8B*	Yes	43.3	49.7	54.2	61.9	51.0	73.8†	86.2	78.5	66.9	67.3	59.4	83.2	45.8	70.6	2412.5	62.3
Embodied-R1-3B*	Yes	51.8	26.5	39.7	69.5	49.4	78.5†	82.7	67.4	47.4	76.3	26.6	—	35.2	—	—	—
Embodied-R1.5-8B*	Yes	61.0	53.8	54.2	74.0	64.8	83.0†	86.9	78.1	69.7	74.7	56.1	—	46.0	—	—	—
Molmo2-ER-4B*	Yes	—	—	52.5	54.0	—	85.7†	87.8	78.8	—	78.0	74.5	—	46.8	—	—	—
Hy-Embodied-VLM-1.0-30B-A3B*	Yes	—	49.6	53.4	65.0	64.6	87.3†	89.7	82.7	69.4	78.0	—	—	60.8	—	—	—
MiMo-Emb-7B*	Yes	62.0	43.0	48.0	63.6	42.35	81.3	88.2	76.2	61.7	78.6	48.5	79.0	46.7	66.3	2320.8	26.4
Thinker-4B*	Yes	62.7	63.7	61.0	72.0	57.4	84.6	86.3	80.2	70.8	72.7	65.4	81.5	—	71.9	2323.4	46.2
Wall-OSS-0.5-3B*	Yes	—	—	—	15.0	—	—	—	—	—	—	—	—	33	44	—	—
Lumo-1-Stage1-7B*	No	—	—	51.0	69.1	—	82.4	86.4	75.6	62.6	74.7	—	—	—	—	—	—
Gemini-ER 2‡	No	—	—	35.4	—	—	90.6†	90.4	81.4	51.1	—	—	—	71.0	—	—	—
Gemini-ER 1.5‡	No	—	—	41.8	48.0	—	—	83.6	73.4	57.7	62.0	39.9	—	47.0	—	—	—
Gemini 2.5 Pro‡	No	—	—	33.6	37.0	—	88.6†	85.9	78.0	71.3	74.7	51.1	—	56.0	—	—	—
Gemini 2.5 Flash‡	No	—	—	41.2	48.0	—	80.3†	85.5	76.2	73.4	73.3	45.3	—	47.5	—	—	—
Gemini 3.1 Pro‡	No	—	—	70.0	61.0	—	86.1†	88.6	—	65.1	—	47.5	—	65.2	—	—	—
GPT-5.6-sol‡	No	—	—	58.3	51.1	—	85.6†	85.2	80.7	66.8	21.3	—	—	64.8	—	—	—
GPT-6-Astra‡	No	—	—	79.6	69.0	—	90.4†	87.3	83.3	73.4	31.3	—	—	77.7	—	—	—

^* Results are sourced from the models' official technical reports or publicly available papers.

^‡ Results were obtained through tests using the models' official APIs.

^† BLINK scores are averaged over the Relative Depth and Spatial Relation subtasks only; all reported results were obtained in our own testing.

Citation

@misc{unifolm-er-1,
  author       = {Unitree},
  title        = {UnifoLM-WLA-1.0: One Model Driven, Whole-Body Coordination},
  year         = {2026},
}