📢 News
- 2026/10/28 [RynnBot Models] We presented several physical AI foundation models in the past few days!
- 2026/10/28 [RynnWorld-Latent] We released RynnWorld-Latent, which learns cross-embodiment world dynamics via 3D-grounded latent actions, enabling joint pretraining on diverse video sources and synthetic data generation for improved robot policy performance. See Project page for more details!
- 2026/10/27 [RynnVLA-Latent] We released RynnVLA-Latent, which combines residual feature prediction, camera-compensated 3D-flow supervision, and camera-motion decoupling to construct continuous transition representations without native robot-action labels. See Project page for more details!
- 2026/10/26 [RynnUnify] We released RynnUnify, a 16B unified embodied model that integrates visual planning, physical reasoning, and robot control, achieving state-of-the-art results on long-horizon embodied tasks! See Project page for more details!
- 2025/10/02 [TCSVT] Sparse-Tuning, an efficient parameter-efficient fine-tuning framework for ViTs, got accepted for IEEE Transactions on Circuits and Systems for Video Technology! Code has been available!
- 2026/09/30 [Preprint] We released RoboChrono, a real robot benchmark for streaming task understanding, evaluating seven tasks grouped into recognition, alignment, and temporal grounding, covering action understanding and anticipation, visual correspondence, temporal ordering, and action localization!
- 2026/09/25 [NeurIPS’26] Two papers, RynnVLA-002 (the upgraded version of WorldVLA) and VLA-Corrector, got accepted for NeurIPS 2026!
- 2026/09/20 [TMLR] MAPGD, a multi-agent framework for black-box prompt optimization, got accepted for TMLR! See Github for more details!
- 2026/07/27 [RynnValue] We presented RynnValue, a scalable value foundation model that predicts temporal distance to goals directly from observations and language instructions. Without relying on manual preference or progress labels, it serves as a zero-shot generalist value function to provide scalable rewards for policy learning across diverse robots, tasks, and viewpoints. See Project page for more details!
- 2026/07/20 [RynnBrain 1.1] We presented RynnBrain 1.1, a systematic upgrade of RynnBrain for embodied intelligence. RynnBrain 1.1 is released in three scales: 2B, 9B, and 122B-A10B. See Project page for more details and amazing videos!
- 2026/07/10 [ACMMM’26] MMaDA-VLA, a fully native pre-trained large diffusion VLA model that unifies multi-modal understanding and generation in a single framework, got accepted for ACMMM 2026! See Project page for more details!
- 2026/07/08 [RynnWorld] We presented RynnWorld-4D and RynnWorld-Teleop!
- RynnWorld-4D is a 4D embodied world model that co-produces future RGB frames, depth maps, and optical flow from a single RGB-D image and a language instruction within one unified diffusion process. We further introduce RynnWorld-4D-Policy, an inverse dynamics head that consumes the internal 4D representations of RynnWorld-4D in a single forward pass, bypassing costly multi-step denoising, to output robot actions in a closed-loop manner. This positions RynnWorld-4D as a powerful 4D action planner for next-generation robotic learning. See Project page for more details! Got #1 Paper of the day on huggingface papers!
- RynnWorld-Teleop is an action-conditioned world model for digital teleoperation, replacing the physical robot to decouple data collection from real-world constraints. Policies trained solely on its generated data achieve zero-shot Sim2Real transfer across diverse bimanual tasks, and augmenting real-world datasets with our digital teleoperation data consistently improves success rates. This positions RynnWorld-Teleop as a high-fidelity, scalable data engine for next-generation robotic learning. See Project page for more details! Got #3 Paper of the day on huggingface papers!
- 2026/06/18 [ECCV’26] Articulat3D, a novel framework for constructing high-fidelity digital twins of articulated objects from casually captured monocular videos, got accepted for ECCV 2026! See Project page for the overview video!
- 2026/06/02 [Preprint] We released STaR-KV, a training-free KV cache compression framework for GUI agents!
- 2026/02/21 [CVPR’26] 2 papers (HiF-VLA and V²Drop) got accepted for CVPR 2026 (Main Conference)!
- 2026/02/18 [RA-L] RoboSimGS, a novel Real2Sim2Real framework that converts multi-view real-world images into scalable, high-fidelity, and physically interactive simulation environments for robotic manipulation, got accepted for RA-L! See Project page for the overview video!
- 2026/02/10 [RynnBrain] We presented RynnBrain, an embodied foundation model grounded in physical reality, including dense (2B, 8B) and MoE (30B) variants, alongside three specialized models: RynnBrain‑Plan (manipulation planning), RynnBrain‑Nav (navigation), and RynnBrain‑CoP (spatial reasoning). See Github and Chinese report from 机器之心. 2026/02/19 We released the technical report!
- 2026/01/31 [ICRA’26] RynnVLA-001, the VLA foundation model, got accepted for ICRA 2026!
- 2026/01/22 [Talk] I gave a talk titled Physical AI Ecosystem: Tackling the Key Barriers to Embodied Intelligence in AAAI-26 Interactive Industry Sessions.
- 2025/12/11 [Preprint] We released HiF-VLA (Hindsight, Insight, and Foresight for VLAs), a unified framework that encodes past dynamics through hindsight priors, anticipates future motion via foresight reasoning, and integrates both through a hindsight-modulated joint expert to enable a “think-while-acting” paradigm for long-horizon manipulation! Project page and Code are available! 2026/02/21 HiF-VLA got accepted for CVPR 2026!
- 2025/11/24 [Preprint] We released RynnVLA-002, an upgraded version of WorldVLA, a more powerful VLA and world model unified model! Get videos and code at Github!
- 2025/11/08 [AAAI’26] 4 papers got accepted for AAAI 2026! They included training-free MLLM inference acceleration methods FiCoCo and GlobalCom2, dexterous grasping policy AffordDex, and tiny-scale VLA VLA-Adapter.
- 2025/09/19 [NeurIPS’25] SSR got accepted for NeurIPS 2025! The work transforms raw depth data into structured, interpretable textual CoT, enhancing spatial reasoning capabilities of MLLMs. See Project page and Github!
- 2025/09/12 [Preprint] We released VLA-Adapter, which reduces reliance on large-scale VLMs and extensive pre-training by using a lightweight Policy module with Bridge Attention, achieving SOTA performance and fast inference speed with minimal computational resources! Checkpoint has been available! See Project page for more details. Got #1 Paper of the day on huggingface papers! 2025/11/08 VLA-Adapter got accepted for AAAI 2026 Oral!
- 2025/08/13 [Preprint] We released AffordDex, a universal grasping policy for dexterous hands with an inherent understanding of both motion priors and object affordances! Grasping videos can be found in Project page! 2025/11/08 AffordDex got accepted for AAAI 2026!
- 2025/08/08 [DAMO RynnBot] We open-sourced RynnEC: a video MLLM for embodied cognition tasks, RynnVLA-001: a VLA model based on pretrained video generation model, RynnRCP: a complete set of robot service agreements and frameworks! 2025/08/11 We released the technical blog for RynnVLA-001! 2025/09/19 We released the technical report for RynnVLA-001!
- 2025/08/02 [CoRL’25] Long-VLA, a novel framework designed to enhance VLA models for challenging long-horizon robotic manipulation tasks, got accepted for CoRL 2025!
- 2025/07/24 [DAMO RynnBot] We released RynnBot PlayGround Beta, a platform that provides data management, SOTA VLA models, model training and validation, cloud-edge collaborative deployment, and so on! Welcome to follow our continuous progress!
- 2025/06/27 [Preprint] We released WorldVLA, an autoregressive action world model that unifies action and image understanding and generation! Code has been available!
- 2025/06/26 [ICCV’25] CARP, Coarse-to-fine AutoRegressive Prediction for visuomotor policy learning, got accepted for ICCV 2025! The approach produces highly accurate and smooth robot actions, achieving up to a 10% improvement of success rates, and delivers 10x faster inference compared to state-of-the-art policies. Paper, code and cool videos can be found in Project page!
- 2025/05/22 [Preprint] We released VARD, a novel RL fine-tuning method on diffusion-based generative models for both protein structure and text-to-image synthesis, enhancing sample quality with improved efficiency, effective mitigation of reward hacking, and broad applicability.
- 2025/05/07 [Preprint] We released OpenHelix, a low-cost open-source dual-system VLA with systematic empirical evaluations on the core design elements. Code and List of papers have been available!
- 2025/03/31 [Preprint] We released Unicorn to explore the question: can high-quality multimodal training data be synthesized purely from text?
- 2025/03/28 [Survey Preprint] We released Exploring the Evolution of Physics Cognition in Video Generation: A Survey, which dives deep into the development of physics cognition in video generation, from basic perception to active cognition! List of papers has been available!
- 2025/03/11 [TCSVT] M2IST, a novel Multi-Modal Interactive Side-Tuning method that effectively addresses the challenges of insufficient multi-modal interaction and high GPU memory consumption, got accepted for IEEE Transactions on Circuits and Systems for Video Technology! Code has been available!
- 2025/02/24 [Preprint] We released Humanoid-VLA, a novel framework that integrates language understanding, egocentric scene perception, and motion control, enabling universal humanoid control!
- 2025/01/28 [ICRA’25] QUART-Online, a novel latency-free quadruped MLLM model that achieves real-time inference while boosting the success rate across various tasks by 65%, got accepted for ICRA 2025! See Project page.
- 2025/01/23 [ICLR’25] ToCa, a token-wise feature caching method that achieves a 2x acceleration for PixArt-α, OpenSora, and DiT while maintaining nearly lossless generation quality, got accepted for ICLR 2025! Code has been available!
- 2025/01/10 [Preprint] We released GlobalCom2, a “global-to-local” approach for training-free acceleration of high-resolution MLLMs with AnyRes strategy. Code has been available! 2025/11/08 GlobalCom2 got accepted for AAAI 2026!