Embodied-R1.5

Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models

Embodied-R1.5 Team

Yifu Yuan^♦^,✻, Yaoting Huang, Xianze Yao, Shuoheng Zhang, Linqi Han, Yutong Li, Pengyi Li, Jiangeng Sun, Wenting Jia, Yucheng Hu, Yuhao Liu, Ruihao Liao, Qiyu Wu, Yuxiao Li, Zhao Zhang, Zibin Dong, Fei Ni, Yan Zheng, Shuyang Gu^✻, Yi Ma^♦^,✻, Hongyao Tang^♦^,✻, Han Hu, Jianye Hao^✻

♦ Project Leader ✻ Corresponding Author

yuanyf@tju.edu.cn

Embodied Foundation Model General Physical Intelligence Unified Embodied Reasoning

Github Paper Models Datasets EmbodiedEvalKit Video

Highlight

TL;DR Embodied-R1 ↗ Embodied-R1.5

Embodied-R1.5 is a unified Embodied Foundation Model (EFM) that integrates comprehensive embodied reasoning capabilities — spanning embodied cognition, task planning, correction, and pointing — within a single 8B-parameter architecture toward general physical intelligence. It achieves SOTA on 16 out of 24 embodied VLM benchmarks, surpassing Gemini-Robotics-ER-1.5 and GPT-5.4, and can be fine-tuned into a VLA that outperforms π₀.₅ across 4 popular manipulation benchmark suites.

Key Contributions

Unified Embodied Foundation Model

We formalize the capability requirements of an EFM and demonstrate that embodied cognition, task planning and correction, and pointing and location can coexist and reinforce each other within a single 8B-parameter architecture, eliminating the fragmented multi-model paradigm.

Large-Scale Data System & Multi-Task Balanced RL

Leveraging three automated data construction pipelines, we build a large-scale data system of over 15B tokens, and design a multi-task balanced RL recipe with output-type-specific rewards that resolves interference inherent in heterogeneous multi-task training.

Planner-Grounder-Corrector Closed-Loop Framework

A Planner-Grounder-Corrector (PGC) closed-loop framework where a single model orchestrates the full autonomy stack — from high-level task decomposition and precise spatial grounding to autonomous self-correction over long-horizon tasks.

SOTA Results & Full Open-Source Ecosystem

Achieves SOTA on 16/24 embodied VLM benchmarks, surpassing Gemini-Robotics-ER-1.5 and GPT-5.4. Fine-tuned as a VLA, it outperforms π₀.₅ across 4 manipulation benchmark suites. We fully open-source model weights, datasets, training code, and EmbodiedEvalKit.

Embodied VLM Evaluation

Per-benchmark scores across all evaluated tasks.

VLA Benchmark Comparisons

Embodied-R1.5-VLA is built upon Embodied-R1.5 as the backbone, augmented with an action head for VLA training — without large-scale action pretraining. We demonstrate that broad improvements in embodied capabilities effectively transfer to downstream VLA manipulation performance.

Models

Closed-Loop Autonomy Framework for Long-horizon Tasks

Embodied-R1.5 integrates Planner, Grounder, and Corrector into a unified closed-loop pipeline, where a single model orchestrates the full autonomy stack — from high-level task decomposition and precise spatial grounding to autonomous reflection and self-correction. This enables complex, long-horizon embodied tasks across diverse robotic embodiments without human intervention.

Core Result

Compared with its predecessor, Embodied-R1.5 features comprehensive advancements in cognition, planning, and error correction, while significantly refining its pointing and localization precision. It currently establishes a new SOTA across 10+ benchmarks, with its mean performance markedly exceeding that of leading general-purpose and embodied models. Notably, Embodied-R1.5 demonstrates a clear competitive advantage over Gemini-Robotics-ER-1.5 across all evaluated capability domains.

Embodied VLM Leaderboard

Comprehensive evaluation across Embodied Planning and Correction (4 Tasks), Embodied Pointing and Location (9 Tasks), and Embodied Cognition and Spatial Reasoning (8 Tasks).

Visual trace prediction evaluation across ShareRobot-V (RMSE↓, DFD↓), VABench-V (RMSE↓, DFD↓), and PIO-S3 (GPT-Score↑).

General Vision Benchmarks

Embodied-R1.5 maintains strong general vision capabilities while acquiring comprehensive embodied skills, demonstrating minimal performance trade-off compared to its base model Qwen3-VL-8B.

Demonstrating superior generalization performance, Embodied-R1.5-VLA excels in all three Simpler evaluation suites. In the rigorous Google Robot Benchmark, our model outperforms the leading baseline, pi0, by a substantial 20% margin, setting a new performance standard for embodied agents.

Simpler-Google Robot Benchmark (Visual Matching)

Model	Pick Coke Can	Move Near	Open/Close Drawer	Open Top Drawer & Place Apple	Overall
RT-1-X	56.7	31.7	59.7	21.3	42.4
RT-2-X	78.7	77.9	25.0	3.7	46.3
OpenVLA	16.3	46.2	35.6	0.0	24.5
SpatialVLA	86.0	77.9	57.4	0.0	55.3
pi0	97.9	78.7	62.3	46.6	71.4
pi0-FAST	75.3	67.5	42.9	0.0	46.4
GR00T N1.5	51.7	54.0	27.8	7.4	35.2
Embodied-R1.5-VLA	92.3	93.8	86.1	97.2	92.4

Simpler-Google Robot Benchmark (Variant Aggregation)

Model	Pick Coke Can	Move Near	Open/Close Drawer	Open Top Drawer & Place Apple	Overall
RT-1-X	49.0	32.3	29.4	10.1	30.2
RT-2-X	82.3	79.2	35.3	20.6	54.4
OpenVLA	54.5	47.7	17.7	0.0	30.0
SpatialVLA	88.0	72.7	41.8	6.3	52.2
pi0	90.1	80.7	27.6	20.5	54.7
pi0-FAST	77.6	68.2	31.3	0.0	44.3
GR00T N1.5	69.3	68.7	35.8	4.0	44.5
Embodied-R1.5-VLA	80.6	72.2	58.3	75.0	71.5

Simpler-WidowX Benchmark (Visual Matching)

Model	Put Spoon on Towel	Put Carrot on Plate	Stack Blocks	Put Eggplant in Basket	Overall
RT-1-X	0.0	4.2	0.0	0.0	1.1
CogACT	71.7	50.8	15.0	67.5	51.3
OpenVLA	4.2	0.0	0.0	12.5	4.2
SpatialVLA	16.7	25.0	29.2	100.0	42.7
pi0	29.1	0.0	16.6	62.5	27.1
pi0-FAST	29.1	21.9	10.8	66.6	32.1
GR00T N1.5	75.3	54.3	57.0	61.3	62.0
Embodied-R1.5-VLA	83.3	75.0	37.5	100.0	74.0

Efficient VLM-to-VLA Adaptation: Embodied-R1.5-VLA achieves high-performance control via direct fine-tuning from a VLM, successfully bypassing the conventional requirement for massive real-robot action pre-training.

Top-Tier Benchmark Dominance: The model demonstrates comprehensive task mastery across all LIBERO suites, reaching first-tier performance and significantly outperforming existing methods that lack action pre-training.

Superior Architectural Stability: Compared to native backbones like Qwen3-VL, Embodied-R1.5 delivers consistent and stable performance gains throughout the entire training curriculum.

LIBERO Benchmark (40 Tasks)

Model	Pt.	Goal	Spatial	Object	Long	Overall
W/ Action Pretraining
SpatialVLA	Y	78.6	88.2	89.9	55.5	78.1
CoT-VLA	Y	87.6	87.5	91.6	69.0	83.9
GR00T N1	Y	93.0	94.4	97.6	90.6	93.9
GR00T N1.6	Y	97.5	97.7	98.5	94.4	97.0
OpenVLA	Y	79.2	84.7	88.4	53.7	76.5
OpenVLA-OFT	Y	97.9	97.6	98.4	94.5	97.1
pi0	Y	95.8	96.8	98.8	85.2	94.2
pi0-FAST	Y	88.6	96.4	96.8	60.2	85.5
pi0.5	Y	98.0	98.8	98.2	92.4	96.9
W/O Action Pretraining
Diffusion Policy	N	68.3	78.3	92.5	50.5	72.4
OpenVLA-OFT	N	91.7	94.3	95.2	86.5	91.9
pi0-FAST	N	89.0	87.0	63.0	48.0	71.8
pi0.5	N	94.6	96.6	97.2	85.8	93.6
Embodied-R1.5-VLA	N	97.4	97.8	99.2	93.2	97.3

LIBERO-Plus Robustness Benchmark

Evaluating model robustness across 7 perturbation types: Camera, Robot, Language, Light, Background, Noise, and Layout changes.

VLM Backbone Comparison on LIBERO

Overall success rates across training steps with different VLM backbones and action experts.

ManiSkill PartNet-Mobility — Affordance Prediction

Model	Seen Categories
Model	Safe	Door	Display	Fridge	Laptop	Lighter	Micro.	Mouse	Box	Trash	Pot	Suitcase	Pliers	Storage	Remote
Where2Act	0.26	0.36	0.19	0.27	0.23	0.11	0.15	0.47	0.14	0.24	0.13	0.12	0.56	0.68	0.07
FlowBot3D	0.67	0.55	0.20	0.32	0.27	0.31	0.61	0.68	0.15	0.28	0.36	0.18	0.21	0.70	0.18
Implicit3D	0.53	0.58	0.35	0.55	0.28	0.66	0.58	0.51	0.52	0.57	0.45	0.34	0.41	0.54	0.39
ManipLLM	0.68	0.64	0.36	0.77	0.43	0.62	0.65	0.61	0.65	0.52	0.53	0.40	0.64	0.71	0.60
Embodied-R1.5	0.98	0.63	0.60	0.78	0.78	0.63	0.90	0.63	0.94	0.41	0.98	0.28	0.34	0.84	0.50

Model	Seen (cont.)					AVG (Seen)	Unseen Categories										AVG (Unseen)
Model	Bottle	Chair	Toaster	Lamp	Disp.	AVG (Seen)	Toilet	Scissors	Table	Stapler	Kettle	USB	Washer	Oven	Faucet	Phone	AVG (Unseen)
Where2Act	0.40	0.13	0.18	0.13	0.40	0.26	0.18	0.35	0.38	0.28	0.05	0.21	0.17	0.20	0.15	0.15	0.21
FlowBot3D	0.26	0.17	0.53	0.29	0.42	0.37	0.23	0.10	0.60	0.39	0.27	0.42	0.28	0.51	0.13	0.23	0.32
Implicit3D	0.43	0.27	0.65	0.20	0.33	0.46	0.45	0.17	0.80	0.53	0.15	0.69	0.41	0.31	0.30	0.31	0.41
ManipLLM	0.64	0.41	0.75	0.44	0.67	0.59	0.38	0.22	0.81	0.86	0.38	0.85	0.42	0.83	0.26	0.38	0.54
Embodied-R1.5	0.72	0.90	0.86	0.35	0.98	0.70	0.71	0.31	0.73	0.75	0.84	0.30	0.92	0.92	0.57	0.33	0.64

Safe

Door

Display

Fridge

Laptop

Lighter

Microwave

Mouse

Box

TrashCan

KitchenPot

Suitcase

Pliers

Storage

Remote

Bottle

Chair

Toaster

Lamp

Dispenser

Toilet

Scissors

Table

Stapler

Kettle

USB

Washer

Oven

Faucet

Phone

Embodied-R1.5 predicts interaction affordances across 30 object categories in ManiSkill PartNet-Mobility.

Real-World Experiments

Embodied-R1.5 demonstrates strong real-world manipulation capabilities across 5 diverse tasks, significantly outperforming prior models in both tool affordance understanding and complex multi-step reasoning.

Real-World Manipulation Results

Pick up [X] and put it on the plate
(Pick&Place)

Put the third duck toy from the left on the plate
(Spatial Reasoning)

Door Open
(Articulated Object Manipulation)

Move [X] to the empty space on the right side of the table
(Tool Affordance) #1

Move [X] to the empty space on the right side of the table
(Tool Affordance) #2

Move [X] to the empty space on the right side of the table
(Tool Affordance) #3

Demo Videos

Clean the vase

Open the Drawer

Plug the charger into the socket

Put all the clutter on the table into the box

Put the specified object into the empty space of the table

Stack the cups into three layers

Sweep the garbage on the ground into the dustpan

Make the milk tea

Take out the blue cup and place it on the coaster

EmbodiedEvalKit

A unified evaluation framework for assessing MLLM on embodied intelligence tasks. View on GitHub →

🚀

Multiple Inference Backends

vLLM (tensor parallel & multi-GPU), HuggingFace Transformers, and API — switch backends in one config.

🤖

20+ Models Supported

GPT, Gemini, Qwen, InternVL, Molmo, Magma and more — generalist and embodied-specialist families out of the box.

📊

20+ Embodied Benchmarks

Embodied QA, Spatial Reasoning, Embodied Pointing, Affordance and Location, and Embodied Planning — all in one place.

🗂️

Standardized Dataset Format

All benchmarks reorganized into HuggingFace Parquet — unified data pipeline for reproducible evaluation.

🎯

Unified Pointing Evaluation

One interface for diverse pointing formats, coordinate systems, and model-specific conventions.

🧩

Modular & Decoupled Design

Models, benchmarks, and metrics cleanly separated — easy to extend, swap, or customize independently.

Supported Benchmarks (20+)

ERQA CV-Bench EmbSpatial SAT RoboSpatial RoboVQA VABench-P Where2Place RefSpatial Part-Afford RoboRefit RoboAfford VSI-Bench OpenEQA EgoPlan2 PIOBench PointBench Pixmo Points Cosmos-Reason BLINK RoboFAC VABench-V PIO-S3 VLABench ...

Evaluated Models (20+)

Generalist

GPT Gemini Qwen InternVL Molmo ...

Embodied

Embodied-R1.5 Embodied-R1 RoboBrain Mimo Pelican Magma VeBrain ...

Acknowledgements and Contribution List

Yifu Yuan proposed the methodology and research direction, developed the complete training dataset, and was responsible for the entire pipeline, including algorithm implementation, model training, inference, and experimental analysis. Yifu Yuan also led the design of the EmbodiedEvalKit framework for evaluation and the drafting of the manuscript.

Yifu Yuan, Hongyao Tang, and Yi Ma served as co-project leads, overseeing the overall execution of the project. The corresponding authors are Shuyang Gu, Yi Ma, Hongyao Tang, and Jianye Hao.

The success of the Embodied-R1.5 project is a collective effort of all contributors. Yaoting Huang, Linqi Han, and Jiangeng Sun contributed to model evaluation. Data construction and cleaning were performed by Yaoting Huang, Linqi Han, Pengyi Li, Jiangeng Sun, Wenting Jia, Yucheng Hu, Zhao Zhang, and Yuxiao Li. The real-world robotic platform setup and experiments were conducted by Shuoheng Zhang, Xianze Yao, Pengyi Li, Yuhao Liu, Yutong Li, Ruihao Liao, Qiyu Wu, and Yuxiao Li. Research guidance and supervision were provided by Shuyang Gu, Zibin Dong, Fei Ni, Yan Zheng, Han Hu, and Jianye Hao.

BibTeX

Embodied-R1.5

@article{yuan2026embodiedr15,
  title={Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models},
  author={Yuan, Yifu and Huang, Yaoting and Yao, Xianze and Zhang, Shuoheng and Han, Linqi and Li, Yutong and Li, Pengyi and Sun, Jiangeng and Jia, Wenting and Hu, Yucheng and Liu, Yuhao and Liao, Ruihao and Wu, Qiyu and Li, Yuxiao and Zhang, Zhao and Dong, Zibin and Ni, Fei and Zheng, Yan and Gu, Shuyang and Ma, Yi and Tang, Hongyao and Hu, Han and Hao, Jianye},
  journal={arXiv preprint},
  year={2026}
}

Embodied-R1 ICLR 2026

@article{yuan2025embodied,
  title={Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation},
  author={Yuan, Yifu and Cui, Haiqin and Huang, Yaoting and Chen, Yibin and Ni, Fei and Dong, Zibin and Li, Pengyi and Zheng, Yan and Hao, Jianye},
  journal={ICLR 2026},
  year={2025}
}

From Seeing to Doing ICLR 2026

@article{yuan2025seeing,
  title={From Seeing to Doing: Bridging Reasoning and Decision for Robotic Manipulation},
  author={Yuan, Yifu and Cui, Haiqin and Chen, Yibin and Dong, Zibin and Ni, Fei and Kou, Longxin and Liu, Jinyi and Li, Pengyi and Zheng, Yan and Hao, Jianye},
  journal={ICLR 2026},
  year={2025}
}