Vision-language models (VLMs) can act as closed-loop robot-use agents (RUAs) by writing and executing control code through tool interfaces. However, turning an open-weight VLM into a high-performing embodied agent requires a complete post-training stack: out-of-the-box open VLMs fail completely at interactive tool contracts, strong frontier teachers are accessible only as black-box endpoints without exposed weights or token probabilities, and supervised imitation alone plateaus under rollout distribution shift. We present EmbodyAgent, an end-to-end post-training framework with three tightly coupled pipelines: (1) a black-box data synthesis pipeline that queries a strong teacher under a forced think-then-code contract and filters trajectories by task success; (2) an SFT pipeline that distills the explanation-and-program targets into an open-weight 9B VLM with both vision and language encoders trainable; and (3) an online RL pipeline that optimizes student-generated episodes via group-relative advantage with a runtime-error-penalized outcome reward. All three pipelines share one primitive-free Python execution tool exposing only raw visual observation and seven-dimensional end-effector control, with no action decoder and no interface shift across stages. Under a strictly separated evaluation protocol on ten LIBERO Spatial tasks (disjoint state splits, greedy test-time decoding, training-only safeguards), held-out success on 100 test states rises from 0/100 (base 9B VLM) to 14/100 (after SFT) and 84/100 (after RL), while training rollout success climbs from 8.4% to 87.7% and median model calls per episode drop from 36 to 11. We further identify a self-reinforcing runtime-error mode induced by outcome-based RL on executable code, and establish a failure-category monitoring and sampling-side curation protocol that stabilizes training.
Figures & tables
Figure 1: The EmbodyAgent full-pipeline post-training framework; the photographs are LIBERO Spatial frames. Left: the single output contract shared verbatim across all stages—at every turn the VLM emits a forced explanation et followed by an executable program ct over two raw runtime operations, shown for two turns of the episode below. Center: the three-stage pipeline. Stage 1 (black-box data synthesis) queries Astra Medium on training-split initial states without accessing teacher weights and filters successful think-then-code episodes into 1,475 targets; Stage 2 (multimodal SFT) distills these targets into an open-weight 9B VLM; Stage 3 (online RL) optimizes the student’s own rollouts via group-relative outcome credit. Bottom: a successful held-out episode from the post-trained 9B VLM (RL round 64), pairing each frame with its verbatim explanation and program through grasp, lift, lower, and place.
Figure 2: Stage 3 RL progression across ten LIBERO Spatial tasks. Left: greedy success on the 100 held-out test states after each round alongside temperature-one training rollout success (round 0 is the Stage 2 SFT checkpoint). The purple dashed line indicates the black-box Astra Medium teacher’s collection acceptance rate (32%, validated episodes over valid attempts, averaged over tasks); under RL, student rollout success rises from 8.4% to 87.7%, surpassing the teacher’s collection rate. The shaded band marks rounds 21–33, where runtime-error episodes dominated failures. Right: median model calls per test episode, inflating briefly during the error wave (rounds 31–33) and compressing to 10–14 calls once the wave subsides.
Task
SFT
RL
0
2/10
10/10
1
2/10
9/10
2
2/10
9/10
3
2/10
9/10
4
0/10
4/10
5
0/10
9/10
Table 1: Per-task success on the held-out states (ten per task), greedy decoding, identical protocol for both checkpoints (RL at round 64).
Figure 3: Per-task held-out success across Stage 3 RL rounds (column 0 is the Stage 2 SFT checkpoint). Every task improves over SFT, with nine of ten tasks reaching 8–10/10; task 4 remains the hardest throughout, matching the task that also resisted Stage 1 black-box synthesis longest.
Figure 4: Stage-by-stage progression on Qwen3.5-9B (success versus median per-episode wall-clock time). Out of the box, the base 9B VLM solves 0/100 held-out states due to format failure. Stage 2 SFT lifts the same VLM to 14/100 (117 s/episode), and Stage 3 RL advances it to 84/100 (136 s/episode, simulation included). The dashed line marks the black-box teacher’s 1,887 s of frontier-API inference per episode, paid once offline during Stage 1 synthesis across 1,620 attempts.
Vision-language-action (VLA) policies generate executable robot actions, but plausible proposals can have different consequences for the task. We present EmbodiedSkills, a skill-based agent framework that augments frozen VLA policies with world action models for consequence-aware action selection. The framework defines reusable skills for scene observation, context encoding, candidate generation, consequence prediction, outcome evaluation, candidate comparison, and action execution. Five stages organize their input-output dependencies in a shared feedback loop, accommodating both continuous action chunks and pick-and-place primitives. At each decision point, independently trained world action models predict the latent consequences of a common set of policy proposals. Selection combines their outcome scores and prediction disagreement with a lightweight comparator that estimates the benefit of an alternative from agreement in its predicted effects. The predictive and comparison modules learn from matched-state candidate branches, while the action policy and pretrained observation encoders remain frozen. On MetaWorld MT50 and CLIPort, EmbodiedSkills improves success rates from 76.28% to 78.64% and from 33.94% to 36.83%, respectively. An additional CALVIN evaluation with the frozen FLOWER policy increases five-task-chain success from 77.10% to 77.90%. These results demonstrate the applicability of consequence-aware selection across different policy backends and action representations.
Wei Wang, Wenqiao Zhang, Yutong Lin +14
College of Computer Science and Technology, Zhejiang University · Nanjing University of Aeronautics and Astronautics · Cornell University +3
Physical AI requires models to ground visual and linguistic understanding in real-world environments while accounting for environmental constraints and execution feedback. We introduce MachEmbodied-VLM (ME-VLM), a unified vision-language model with two variants, 4B and 35B-A3B, that brings together embodied cognition and multimodal agent capabilities. Our work emphasizes physical perception and spatiotemporal reasoning, together with planning, interaction, and outcome assessment in both digital and physical environments. We construct training data spanning embodied and multimodal agent tasks, including execution observations and feedback to support outcome assessment and decision refinement. The training pipeline comprises embodied capability injection, separate reinforcement learning of embodied and multimodal-agent experts, and multi-teacher on-policy distillation that consolidates their complementary capabilities into a single model. Experiments show competitive performance on both embodied and agent benchmarks, as well as on autonomous-driving and embodied-navigation tasks. For edge deployment, visual token compression, W4A8 quantization, and hardware-software co-optimization enable on-device inference of the 4B variant on the M100, reducing prefill latency from 400 ms to 188 ms. Project Page: https://machembodied.com/ME-Brain/ME-VLM.html Code Repository: https://github.com/MachEmbodied/ME-VLM
We introduce HY-Embodied-0.5, a family of foundation models specifically designed for real-world embodied agents. To bridge the gap between general Vision-Language Models (VLMs) and the demands of embodied agents, our models are developed to enhance the core capabilities required by embodied intelligence: spatial and temporal visual perception, alongside advanced embodied reasoning for prediction, interaction, and planning. The HY-Embodied-0.5 suite comprises two primary variants: an efficient model with 2B activated parameters designed for edge deployment, and a powerful model with 32B activated parameters targeted for complex reasoning. To support the fine-grained visual perception essential for embodied tasks, we adopt a Mixture-of-Transformers (MoT) architecture to enable modality-specific computing. By incorporating latent tokens, this design effectively enhances the perceptual representation of the models. To improve reasoning capabilities, we introduce an iterative, self-evolving post-training paradigm. Furthermore, we employ on-policy distillation to transfer the advanced capabilities of the large model to the smaller variant, thereby maximizing the performance potential of the compact model. Extensive evaluations across 22 benchmarks, spanning visual perception, spatial reasoning, and embodied understanding, demonstrate the effectiveness of our approach. Our MoT-2B model outperforms similarly sized state-of-the-art models on 16 benchmarks, while the 32B variant achieves performance comparable to frontier models such as Gemini 3.0 Pro. In downstream robot control experiments, we leverage our robust VLM foundation to train an effective Vision-Language-Action (VLA) model, achieving compelling results in real-world physical evaluations. Code and models are open-sourced at https://github.com/Tencent-Hunyuan/HY-Embodied.