Tencent Releases Two Embodied AI Models—Hy‑Embodied‑VLM‑1.0 & RxBrain‑1.0—to Boost Robot Real‑World Understanding

Hy-Embodied-VLM-1.0

Hy-Embodied-VLM-1.0 optimizes the "understand‑act‑adapt" loop for embodied agents by constructing a three‑level capability hierarchy: physical‑state understanding, action‑change reasoning, and temporal‑adaptive inference. The model is built on the Hunyuan A3B base and trained with large‑scale embodied data.

Training uses a staged data pipeline. The mid‑train stage aggregates over 18 million question‑answer pairs that cover state understanding, action transition, and complex reasoning. The post‑train stage adds roughly 48 k high‑quality instruction samples to strengthen counting, spatial relations, interaction understanding, action feasibility, goal localization, planning, and error correction.

On a 37‑task embodied benchmark the model achieves 68.6 in physical‑state understanding, 64.1 in action‑change reasoning, and 57.4 in temporal‑adaptive inference, yielding an average score of 65.6. This performance approaches the previous A32B flagship while using only one‑tenth of the compute and surpasses same‑size models Qwen3.6‑A3B, Cosmos 3‑A8B, and Embodied‑R1.

Hy-Embodied-RxBrain-1.0

Hy-Embodied-RxBrain-1.0 unifies textual reasoning and visual imagination around a single task goal, generating alternating sequences of sub‑task text and target images within a continuous context.

Pre‑training consumes more than 5 × 10⁴ hours of high‑quality embodied data: 31 568 h first‑person/UMI video, 17 292 h real‑robot recordings, and 1 317 h simulation data. Open‑source data account for 57 % (28 597 h). An automated video‑segmentation pipeline converts raw videos into atomic action sequences, producing approximately 2.1 billion training samples across four granularity levels (L0–L3): continuous state imagination, atomic action planning, high‑level sub‑task planning, and final goal‑state imagination.

During mid‑training an additional 35 million samples are introduced, covering spatial reasoning, multi‑view understanding, causal inference, visual localization, behavior planning, 3‑D perception, error analysis, and multimodal generation. Auxiliary visual imagination is injected into the inference process.

RxBrain‑Bench evaluates the model on three stages: Embodied VQA (1 381 samples), World State Prediction (1 116 samples), and Joint Subgoal Planning (3 640 samples). In fully free‑rolling evaluation RxBrain attains a composite planning score of 0.68, outperforming Cosmos3‑Nano (0.521), BAGEL‑7B‑MoT (0.503), and a modular Qwen‑Agent (0.431). Detailed metrics show 0.83 for observation understanding, 0.78 for sub‑task planning, and 0.52 for target‑image correctness; performance declines as rollout steps increase.

For short‑term video generation RxBrain scores 0.62, higher than the general video model Wan2.2‑TI2V‑5B (0.429) and comparable to the specialized world model Cosmos3‑Nano (0.591). In the GenEval text‑to‑image benchmark RxBrain achieves 82.4, matching the dedicated model BAGEL (82) and exceeding Cosmos3‑Nano (71.68).

Real‑World Robot Validation

Physical experiments on DOBOT X‑Trainer and Ark Unlimited A5 arms evaluate three multi‑stage tasks: setting a table, folding and storing glasses, and picking trash. Success rates are 97 % (table), 95 % (fold‑store), and 68 % (trash), yielding an average of 87 %, which exceeds the π0 baseline (68 %) and the π0.5 baseline (82 %).

Open‑Source Availability

Hy‑Embodied‑VLM‑1.0 repository: https://github.com/Tencent-Hunyuan/HY-Embodied

Model on Hugging Face: https://huggingface.co/tencent/Hy-Embodied-VLM-1.0

Hy‑Embodied‑RxBrain‑1.0 repository: https://github.com/Tencent-Hunyuan/Hy-Embodied-RxBrain-1.0

Model on Hugging Face: https://huggingface.co/tencent/Hy-Embodied-RxBrain-1.0