Ant Lingbo Technology has launched LingBot-VA, an autoregressive framework that unifies video world modeling with policy learning, jointly learning future frame prediction and action execution. Evaluation results show that the framework demonstrates clear potential in long-horizon manipulation, data-efficient post-training, and generalization to new scene configurations.
X Square Robot has released X-Tokenizer, a cross-modal embodied action tokenizer built on a semantic residual quantization (SRQ) architecture and pretrained on 17 robotic arm series and 2.4 million trajectories (including 2 billion action frames). Compared with FAST, it improves multimodal alignment by 13.5% and long-horizon task performance by 8.25%, and scores 82.8 on RoboTwin 2.0; in real-robot tests it outperforms comparison methods such as FAST and 4-level RVQ.
Why it matters: The original text presents X-Tokenizer's method design and comparative experimental data, letting readers understand the action tokenizer's multimodal alignment approach and its reusable effects.
HUMANOID has launched KinetIQ Ascend, an end-to-end visual reinforcement learning method that operates under production deployment conditions, extending its KinetIQ framework to let robots improve manipulation skills through trial and error rather than relying solely on human demonstration; the company says this is the first such method publicly demonstrated on production-line humanoid robot manipulation.
Why it matters: The original text provides specific throughput and success-rate figures for three industrial tasks and notes a pattern of predictable performance improvement with training duration, making it useful for assessing the practical value of real-world reinforcement learning on humanoid robots.
Galbot released AstraBrain-WBC 0.5 (paper titled Humanoid-GPT, accepted to CVPR 2026), training a Transformer whole-body controller on about 2 billion frames of retargeted motion data, achieving a 92.58% zero-shot motion tracking success rate with 0.39 ms inference latency; code and data are open-sourced.
Why it matters: This analysis clearly explains the data scale, architecture choices, and limitations of Galbot's new model, letting readers assess progress on the general motion control route.
Our Data Foundation Model (DFM) transforms raw human interactions into high-precision, multimodal, model-ready data. Through ground-truth evaluation and continuous feedback, DFM builds a scalable human data flywheel for physical AI.
Built on our Data Foundation Model (DFM), GenRobot.AI delivers higher-quality, more scalable, and more efficient Ego+Fingers data production. The industry-first DFM enables end-to-end data processing. #EmbodiedAI #HumanData #6DPose #PhysicalAI #RobotLearning #EgoData
Kepler has released the Kepler-OmniVTLA native force-tactile omni-modal large model, paired with the Kepler-OmniTac™ force-tactile full-stack data collection solution, natively compatible with vision, six-axis force, and tactile multi-modal information.
Embodied general intelligence foundation model company DeepWise Robotics announced the completion of an angel round of hundreds of millions of yuan, with investment from Puhua Capital, Oriental Fortune Capital, Zhongguancun Capital and other institutions. The funds will mainly be used to accelerate the closed-loop construction of the "human data flywheel" and advance commercialization in education, elderly care and other scenarios.
Figure AI demonstrated two humanoid robots equipped with Helix-02 collaborating to tidy a bedroom within two minutes, completing tasks such as opening a door, hanging clothes, putting away headphones, closing a book, taking out trash, pushing a chair under a desk, and making a bed together. The two robots run the same learned vision-language-action (VLA) policy, with no shared planner, message passing, or central coordinator; each infers the other's intent from its actions via its own cameras. The company says that to its knowledge this is the first demonstration of multi-humanoid collaborative mobile manipulation completed by a single learned neural network directly from pixels to actions.
Why it matters: The official demonstration shows two humanoid robots driven by the same VLA system collaborating to tidy a bedroom, giving readers insight into an approach to multi-robot collaboration without a shared planner.
Genesis AI has released GENE-26.5, calling it the first "robot brain" on the path toward human-level capability. It is a full-stack approach combining a robot-native foundation model, a 1:1 humanoid robotic hand, non-invasive data-collection gloves that capture motion, force, and touch, and a simulator. The model is trained jointly on language, vision, proprioception, touch, and action, and was tested on a set of self-designed tasks; the demonstration was fully autonomous, at 1x speed, using a single model with the same weights.
Genesis AI's robotic foundation model GENE-26.5 autonomously completed a continuous one-handed fine manipulation sequence: unwrapping a straw, flipping the straw with one hand, dropping it into a cup, and then handing the drink to a person. The entire sequence was executed autonomously and filmed at 1x speed, not teleoperated. The demo showcases fluid one-handed dexterous manipulation across multiple small, precise actions.
Genesis AI's robotic foundation model GENE-26.5 autonomously arranges and bundles wires at 1x speed. Wire harness assembly is regarded as one of the hardest tasks in electronics and electrical engineering, and this demonstration shows the precise, coordinated hand control required to achieve human-level dexterity.
Genesis AI demonstrated a robot autonomously making a smoothie end to end at 1x real speed, including handling ingredients, pouring, blending, and two-hand coordinated mid-air serving. GENE-26.5 is Genesis AI's foundation model for human-level physical manipulation.
On April 10, Lingchu Intelligence released Psi-R2, pretrained on human data, and the Action-Conditioned World Model Psi-W0, and open-sourced the first 1,000 hours of a full-modality human hand manipulation dataset, with total data reserves reaching 100,000 hours.
Why it matters: The original text provides the division of labor between the two new models and the open-source data entry point, so readers can assess the feasibility of using human data for robot training.
Physical Intelligence announced that its latest model, π0.7, has several emergent capabilities: it can control a new robot to fold shirts without any shirt-folding data, can learn to use appliances through language guidance, and can complete multiple dexterous manipulation tasks within a single model.
Physical Intelligence has released the general-purpose robot model π0.7, which it says can perform a variety of dexterous manipulations with speed and robustness comparable to fine-tuned specialized models, and shows preliminary signs of compositional generalization on tasks not seen in its training data.
Why it matters: The article uses two specific experiments, an air fryer and folding clothes with a UR5e, to show the actual performance of compositional generalization and cross-embodiment transfer, making it easier to judge the gap between general-purpose and specialized models.
Lingchu Intelligence released two models, Psi-R2 and Psi-W0, which it calls the industry's first embodied intelligence models pretrained on human data at the 100,000-hour scale.
Why it matters: The original article, written by the team itself, breaks down the dual-model architecture and data governance conclusions, offering a reusable path for human data utilization along with experimental evidence.
Physical AI company Octopus Dynamics (SynapX) announced the completion of a nearly $50 million first funding round. The investors are Horizon Robotics, Hillhouse Venture Capital, Xiaomi Strategic Investment, Shunwei Capital, and Linear Capital. The company was founded in January 2026, and less than two months passed from its founding to the closing of this first round.
Figure AI released a new Helix 02 demo in which a single neural network system completes an end-to-end living room tidying task with full-body control, continuously manipulating objects, tools, and containers while walking. In the demo, Helix 02 uses no new algorithms or special engineering; with only added data, it learned to wipe and clean with a spray bottle and towel, collect blocks into a box carried with both hands, toss a pillow back onto the sofa, turn around while holding a remote and then press a button to turn off the TV, and sidestep through the narrow gap between the coffee table and the sofa.
We developed a memory system for our models that provides both short-term visual memory and long-term semantic memory. Our approach enables us to train robots to perform long and complex tasks, such as cleaning a kitchen or preparing a grilled cheese sandwich from scratch 👇
Dexterity has unveiled its world model Foresight, an intelligence layer that enables robots to reason about the physical world, predict changes, and execute operations; it was trained on real experience from more than 100 million autonomous actions in enterprise logistics production environments.
Titan o1: connecting two worlds in an instant ❤️❤️❤️. With the upgraded GAE model, we have unified dexterous manipulation and whole-body movement as never before. Experience seamless cross-limb integration that makes "out of reach" within reach. Titan o1. Arrive, wherever you are. 🎆🎆🎆
Humanoid has released KinetIQ, a proprietary AI framework for end-to-end unified orchestration of humanoid robot fleets in industrial, service, and home scenarios; a single system can control robots with different body and end-effector forms.
Figure AI has released Helix 02, extending control from the upper body to the entire humanoid robot and enabling continuous, integrated autonomous tasks combining locomotion and manipulation. It adds a new System 0 whole-body controller, trained on over 1,000 hours of human motion data and simulation-based reinforcement learning, which outputs joint commands at 1 kHz via a 10M-parameter network, replacing roughly 109,504 lines of hand-written C++ code; System 1 connects head-mounted vision, palm cameras, fingertip touch sensing, and proprioception, outputting whole-body joint targets at 200 Hz.
Why it matters: The original text provides details on Helix 02's architectural layers, training data, and demonstration tasks, allowing readers to understand the specific implementation approach for full-body autonomous control.
Sharpa has released CraftNet, an end-to-end hierarchical VTLA (vision-tactile-language-action) model aimed at last-millimeter contact control in fine manipulation. The model consists of three parts: System 2 (about 1 Hz, an open-source VLM responsible for reasoning and task decomposition), System 1 (about 10 Hz, responsible for motion planning and coarse actions), and System 0 (about 100 Hz, performing real-time fine-tuning of hand movements based on tactile feedback), with the three cooperating asynchronously.
We got robots to wash pans, clean windows, make peanut butter sandwiches, and more tasks! Fine-tuning our latest model makes all these tasks possible, with interesting implications for robotics, the Moravec paradox, and the future of large models in embodied intelligence. See details below!
Introducing ACT-1: the frontier robot AI behind Memo. ACT-1 enables Memo to perform: 🎭 ⏰ Long-duration tasks, such as tidying up after dinner 🦾 Dexterous tasks, such as folding socks ☕ Everyday enjoyments, such as making espresso How does it work? 🧵
Sunday Robotics released the foundation model ACT-1, claiming it was trained without any robot manipulation data; instead, human motions were captured with a Skill Capture Glove and converted into robot data via Skill Transform with a 90% success rate.
Physical Intelligence has released a new model, π0.6, which can learn from its own experience through reinforcement learning (RL). Compared with a baseline model not trained with RL, π0.6 can more than double throughput and can perform real-world tasks, including making espresso drinks, folding various garments, and assembling cardboard boxes.
Why it matters: The original text provides a comparison showing more than double the throughput of a baseline model not trained with reinforcement learning, as well as the scope of three types of real-world tasks.
Physical Intelligence released the VLA model π*0.6. After training with its Recap (RL with Experience & Corrections via Advantage-conditioned Policies) method, throughput on some of the hardest tasks more than doubled, and the failure rate can be reduced to half of the original or lower.
Why it matters: The original article explains how compounding errors arise in VLA under imitation learning and presents a training approach that uses expert corrections and autonomous experience for reinforcement learning.
We're excited to share our latest progress on DYNA-1 pretraining! 🤖 The foundation model can now perform a variety of dexterous tasks (folding clothes, parcel sorting, ...) without any post-training, even in environments it has never seen. This powerful foundation also supports extremely efficient fine-tuning: with only about 1 hour of data, it can reach about 100% success on challenging new tasks! 🤯 Check out two tasks it has mastered: stacking cups and cutting celery, performed repeatedly with zero failures. 👇
🎃 Halloween is coming. Our hardworking team is lining up for candy, of course, served by Dynasaur! The DYNA VLA model now has robust agentic reasoning capabilities and can serve any combination and quantity of candy! Given that the number of possible combinations is exponential, pure imitation learning will not work. No video editing. Continuous live footage, as always 🤖 DYNA wishes everyone a happy Halloween! 🍬
KEENON Robotics has officially released KOM 2.0 (KEENON Operator Model 2.0), which it calls the world's first VLA (vision-language-action) model for the service industry. The model uses a dual-system architecture with the K-Mind thinking system and the K-Act action system, is trained on hundreds of millions of real service interaction data points, and supports millisecond-level real-time decision-making.
Figure announced two advances in Project Go-Big: a partnership with Brookfield to build a large-scale humanoid pretraining dataset, with Brookfield holding over 100,000 homes, 500 million square feet of commercial office space, and 160 million square feet of logistics space; and Helix, after training on 100% first-person human video, responding to natural-language commands such as "go to the fridge" without any robot demonstrations, navigating cluttered real home environments.
Why it matters: The original gives the training method by which Helix learned navigation using only human video and the scale details of the Brookfield data partnership, offering a view of actual progress on the humanoid pretraining data approach.
Figure AI announced that its VLA (vision-language-action) model Helix learned to load a dishwasher using only new data, without changing the algorithm or adding special engineering. Helix can separate stacked plates and load them in order, pick up a glass with one hand, reorient it, and carefully place it with the other hand; it can also handle cluttered starting states and recover from grasp failures or collisions. The same model previously completed towel-folding and package-sorting tasks, and the company says this reflects the direction of a single general system learning new capabilities incrementally with data.
LimX Dynamics has introduced LimX DreamActor, an embodied intelligence training paradigm that for the first time deeply integrates Real2Sim2Real, imitation learning, and real-robot reinforcement learning into a full pipeline of "real-world environment data collection—simulation data pretraining—real-robot data post-training." The system can complete multi-view data collection using only a consumer-grade camera; through 3D reconstruction and physical parameter alignment it builds a highly realistic simulation scene, then uses a small amount of real-robot reinforcement learning to complete cross-domain transfer, aiming to improve sample efficiency, convergence speed, and deployment stability.
Figure AI's VLA model Helix fully autonomously folds towels, which the company says is the first time a multi-fingered humanoid robot has completed this task using an end-to-end neural network. With the architecture and training hyperparameters unchanged and only a new dataset added, it can pick items from a mixed towel pile, adjust its folding strategy based on the initial shape, correct multiple grasping errors, and untangle twisted towels using fine motions such as tracing edges with its thumb and pinching corners.
Why it matters: The same model completes the towel-folding task with no architecture changes, relying only on added data, demonstrating a generalization path from logistics to deformable object manipulation.
Figure AI released progress on Helix's capabilities after three months of deployment in a logistics environment: per-item processing time fell from about 5.0 seconds to 4.05 seconds, barcode scanning success rate rose from about 70% to about 95%, and the system gained the ability to handle flexible packages such as poly bags and flat envelopes.
Why it matters: The original article provides the data scale and ablation experiment details for each architecture module, allowing readers to understand the specific sources of the logistics performance improvement.
This article explains NVIDIA's open-source humanoid robot foundation model Isaac GR00T N1, unveiled at the March 2025 GTC conference, along with the accompanying GR00T Blueprint synthetic data toolchain and Newton physics engine.
Figure AI released an end-to-end reinforcement learning walking controller that lets the Figure 02 humanoid robot learn a human-like gait in simulation by imitating human reference trajectories through rewards, including heel strike, toe push-off, and arm swing. Training runs in parallel across thousands of robots in GPU-accelerated physics simulation, collecting years of simulated data in hours, and combines domain randomization with kHz-level torque feedback control to achieve zero-shot transfer to the real robot; 10 Figure 02 units run the same neural network without individual tuning.
Why it matters: The original text provides details on the training method and sim-to-real transfer for Figure 02's reinforcement-learning-based human-like walking, offering insight into its technical approach.