# g1 Published articles for g1. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## STRIDE: Learning Humanoid Locomotion in Minutes with Acceleration-Based Rewards DevFeed: [STRIDE: Learning Humanoid Locomotion in Minutes with Acceleration-Based Rewards]() Original publisher: [Read original article]() Author: Nico Bohlinger, Jan Peters Published: 2026-11-09T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Reinforcement learning](), [Humanoid Robots](), [Machine learning](), [data augmentation]() Tags: [acceleration](), [algorithm](), [cs-lg](), [cs-ro](), [falls](), [g1](), [gap](), [humanoid](), [imu](), [learning](), [minutes](), [motion-capture](), [reinforcement-learning](), [reward](), [rl](), [robot](), [simulation](), [symmetry](), [training]() ### AI overview STRIDE trains humanoid locomotion using a reward based on following commanded acceleration, which requires only an onboard IMU rather than base-velocity measurements from motion capture or a well-calibrated state estimator. The method combines off-policy RL algorithm XQC with high-gain joint position control and left-right symmetry augmentation. In simulation, it learned to walk on Unitree G1 and Booster T1 robots within 14 minutes of samples. With the two added components, training falls decreased by 81%; without them, learning took over an hour. The supplied text is an abstract and does not establish results on physical robots. ### Source excerpt arXiv:2611.08857v1 Announce Type: cross Abstract: Learning locomotion directly on a humanoid robot is a promising direction to avoid the sim-to-real gap. However, locomotion rewards in Reinforcement Learning (RL) typically track the base velocity, which cannot be measured onboard without an external motion capture system or well-calibrated state estimator. This is why we introduce \method{}, which learns humanoid locomotion from minutes of samples with a reward that only requires the onboard Inertial Measurement Unit (IMU). Instead of tracking the base velocity, STRIDE rewards the robot for following a commanded acceleration profile. To learn efficiently from this reward, we use the off-policy RL algorithm XQC and combine a high gain setting for the joint position controllers with left-right symmetry augmentation. In simulation, STRIDE learns to walk on the Unitree G1 and Booster T1 humanoids within 14 minutes of samples. Compared to learning without these two components, it reduces the falls during training by 81%, while without them the robot learns only in over an hour. ## Social Humanoid: Socially Responsive Human-Humanoid Interaction DevFeed: [Social Humanoid: Socially Responsive Human-Humanoid Interaction]() Original publisher: [Read original article]() Author: Vongani H. Maluleke, James Ni, Kie Horiuchi, Lea Wilken, Jiahui Lei, Evonne Ng, Roei Herzig, Jitendra Malik, Trevor Darrell, Angjoo Kanazawa Published: 2026-11-06T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Humanoid Robots]() Tags: [abstract](), [adaptive](), [agents](), [apply](), [arxiv](), [behavior](), [benchmark](), [body](), [bridge](), [capture](), [controller](), [cs-cv](), [cs-lg](), [cs-ro](), [datasets](), [flexible](), [g1](), [gap](), [https](), [humanoid](), [interaction](), [responsive](), [social]() ### AI overview Social Humanoid generates whole-body motion for a 29-DoF Unitree G1 from a human partner's motion, with the stated aim of coordinating more naturally during physical interaction. Its high-level policy models human and robot motion using a shared tokenized representation, reacts to the partner, and predicts their motion when they are occluded; a low-level controller tracks the generated motion on the robot. For training data, a two-phase adaptive retargeting method maps one partner to the robot and morphs the other to preserve contact dynamics, reachability, and relative body scale. Applied to five motion-capture datasets, this produced a 260-hour human-humanoid interaction benchmark. The authors report coordinated, human-aligned behavior across varied interactions in simulation and real-world trials. ### Source excerpt arXiv:2611.08407v1 Announce Type: cross Abstract: Humanoid robots are typically trained as solitary agents, leaving them unable to coordinate physically with human partners who naturally anticipate and adjust to one another. To bridge this gap, we present Social Humanoid, a hierarchical system that generates and executes socially responsive whole-body motion on a 29-DoF Unitree G1 conditioned on a human partner's motion. Its high-level policy performs flexible sequence modeling of human and humanoid motion through a shared tokenized representation, generating reactive humanoid motion and anticipating the partner's motion when they are occluded, while a low-level whole-body tracking controller executes the generated motion on the robot. To train the system, we introduce a two-phase adaptive retargeting procedure that retargets one interaction partner to the robot while morphing the other partner to preserve contact dynamics, reachability, and relative body scale. We apply this procedure across five motion-capture datasets to produce a new Human-Humanoid Interaction (HHI) benchmark of 260 hours of human-humanoid interactions. Our results demonstrate that Social Humanoid produces coordinated, human-aligned behavior across diverse interaction types in simulation and the real world. Project page: https://von31.github.io/social-humanoid/ ## $\\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation DevFeed: [$\\nabla$Mimic: Ultra-Efficient Humanoid Control with Differentiable Simulation]() Original publisher: [Read original article]() Author: Weyl Lu, Yubei Chen Published: 2026-11-02T05:00:00Z Content type: article Language: en Sources: [arXiv: Machine Learning]() Topics: [Humanoid Robots](), [Robotics](), [Reinforcement learning](), [World models]() Tags: [clipping](), [control](), [cs-lg](), [cs-ro](), [engine](), [g1](), [gpu](), [gradient](), [gradient-clipping](), [humanoid](), [infrastructure](), [kernels](), [learning](), [median](), [model](), [motion](), [motion-tracking](), [open-source](), [open-source-infrastructure](), [policy](), [ppo](), [simulation]() ### AI overview $\nabla$Mimic uses exact simulator gradients for humanoid motion tracking, stabilizing rare unbounded gradient samples with median-relative clipping at each backward step and reusing simulated windows through multistep regression to bounded action targets. NablaEngine accelerates the MuJoCo contact-dynamics adjoint for the Unitree G1 with fused GPU kernels; its analytic gradients match float64 finite differences and run 11.5x faster than the available MuJoCo Warp adjoint. On all 40 LAFAN1 clips with one policy on one GPU, the method reaches score 0.82 in 20 minutes, matching the cited world-model result reached in 1.2 hours, and scores 0.874 in 1.8 hours. The abstract reports that the first-order baseline SHAC stops within minutes due to exploding gradients. ### Source excerpt arXiv:2610.41654v1 Announce Type: cross Abstract: Robot policy learning is dominated by zeroth-order methods such as PPO, which treat the simulator as a black box and pay for it with enormous sample counts. First-order policy learning differentiates through the dynamics, but on contact-rich whole-body control it has rarely worked: its gradients explode. A recent world-model method showed that derivative information can make humanoid motion tracking far more efficient, using surrogate gradients of a learned model. We show that exact gradients of the simulator itself do better, once two obstacles are removed. Exact gradients contain rare unbounded samples, because contact dynamics is not differentiable everywhere and its derivative diverges near some events; and they are expensive, because the simulator's adjoint has to replay the contact solve. $\nabla$Mimic bounds these samples with median-relative per-sample gradient clipping at every step of the backward pass, and reuses each costly simulated window by multistep regression onto gradient-derived action targets, which bounds every action target independently of the gradient scale. NablaEngine, a differentiable GPU simulation engine for the Unitree G1, removes the cost: it computes MuJoCo's full contact dynamics with fused per-world kernels and the adjoint of a whole control step as a single megakernel, with analytic gradients that match float64 finite differences, $11.5\times$ faster than the available MuJoCo Warp adjoint. On whole-body tracking of all 40 LAFAN1 clips with one policy on one GPU, $\nabla$Mimic reaches the score reported by the state-of-the-art world-model method, 0.82, in 20 minutes instead of 1.2 h, and 0.874 in 1.8 h. SHAC, the standard first-order baseline, stops on exploding gradients within minutes. We release NablaEngine and $\nabla$Mimic as open-source infrastructure for first-order humanoid policy learning. ## METATRON: A Promptable Whole-Body Execution Layer for Humanoids DevFeed: [METATRON: A Promptable Whole-Body Execution Layer for Humanoids]() Original publisher: [Read original article]() Author: Chen Tessler, Yifeng Jiang, Alec Otto, Yi Shi, Umar Iqbal, Gal Chechik, Xue Bin Peng Published: 2026-10-30T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Humanoid Robots](), [Loop Engineering]() Tags: [balance](), [code](), [coding-agents](), [commands](), [cs-ai](), [cs-ro](), [evaluations](), [furniture](), [g1](), [generative](), [goals](), [humanoids](), [motion-capture](), [pipelines](), [policy](), [prompt](), [robustness](), [skills](), [stateful](), [style](), [synthetic](), [unitree]() ### AI overview Metatron pairs a frozen generative whole-body policy with stateful skills written as ordinary code. Skills issue timed body-part targets, velocity commands, and style cues at each control step; the policy turns these partial prompts into coordinated motion. The policy is trained using motion capture, synthetic recovery trajectories, and retargeted scene interactions, while coding agents refine skill rules from motion data without task-specific policy training. On a physical Unitree G1, one checkpoint supports walking, recovery, teleoperated object pickup, and sitting. The reported comparisons show better reaching success and push robustness than the tested reference-tracking pipelines. Climbing and acrobatics are demonstrated on a simulated character. ### Source excerpt arXiv:2610.38027v1 Announce Type: cross Abstract: Humanoid controllers often require task-specific training despite sharing the need for balance and whole-body coordination. We present Metatron, a promptable whole-body execution layer combining a generative base policy with a programmable skill layer. The base policy receives partial goals, collectively called a prompt, that can combine timed body-part targets, velocity commands, and style cues. It produces diverse whole-body motions consistent with these goals. Training combines motion capture, synthetic recovery trajectories, and retargeted scene-interaction motions. After training, the policy is frozen. Stateful programs called skills express task intent as prompts at each control step, drawing rough geometric sketches that the policy turns into coordinated movement. Skills are ordinary code, refined by coding agents that analyze motion data and fit compact rules without task-specific policy training. On the physical Unitree G1, one frozen checkpoint supports walking, recovery, teleoperated object pickup, and sitting on furniture. Comparative evaluations show higher reaching success and push robustness than the tested reference-tracking pipelines. The same recipe produces climbing and acrobatics on a simulated character. Metatron enables coding agents to turn learned motor capabilities into new skills through code, without further policy training. ## JDK 27 G1/Parallel/Serial GC changes DevFeed: [JDK 27 G1/Parallel/Serial GC changes]() Original publisher: [Read original article]() Author: Thomas Schatzl Published: 2026-10-30T00:00:00Z Content type: article Language: en Sources: [Inside Java]() Topics: [openjdk](), [JDK Flight Recorder](), [JEP-401]() Tags: [g1](), [garbage-collectors](), [gc](), [jdk-27](), [parallel](), [serial]() ### AI overview JDK 27 makes G1 the default garbage collector across environments; Serial GC remains selectable with `-XX:+UseSerialGC`, and the excerpt notes that G1 is not best for every application. G1 changes also remove heap resizing after Full GC based on the default MinHeapFreeRatio and MaxHeapFreeRatio values, improve adaptive concurrent-marking behavior, and fix reclaiming some humongous objects held alive by weak references. Parallel GC can now lower its adaptive tenuring threshold and expand the heap after repeated large allocations cause Full GCs. Across collectors, TLAB sizing better handles many short-lived, lightly allocating threads, and string-deduplication logs and JFR events improve. The excerpt says there were no notable Serial-specific changes beyond its reduced default use; about half of the roughly 350 GC changes were refactorings and cleanups, and another 35% were fixes and robustness work. ### Source excerpt Now that the JDK 27 has been released, I thought it would be time to discuss interesting changes in OpenJDK's stop-the-world collectors in the HotSpot VM once more. ## REACH: An RL-based Hierarchical Humanoid Crawling Framework for Adaptable Confined Space Traversal DevFeed: [REACH: An RL-based Hierarchical Humanoid Crawling Framework for Adaptable Confined Space Traversal]() Original publisher: [Read original article]() Author: Shengzhang Xu, Yuning Guo, Weiyi Hu, Yifeng Xu, Zhongyu Li, He Kong Published: 2026-10-28T04:00:00Z Content type: article Language: en Sources: [arXiv: Artificial Intelligence]() Topics: [Humanoid Robots]() Tags: [context](), [cs-ai](), [cs-ro](), [experiments](), [framework](), [g1](), [history](), [humanoid](), [interaction](), [lidar](), [planner](), [point](), [policies](), [proactive](), [real-world](), [rl](), [space](), [tracking](), [traversal](), [unitree](), [variable](), [velocity]() ### AI overview REACH enables a Unitree G1 humanoid to traverse height-constrained spaces by combining variable-height bipedal walking (1.27-0.50 m) with crawling (0.50-0.20 m). A history-conditioned latent representation uses proprioceptive observations to capture terrain and contact context, while a bidirectional policy connects the two modes across their kinematic and contact discontinuities. A LiDAR planner turns point clouds into collision-free 3D paths and locomotion commands with proactive height targets. The abstract reports real-world demonstrations of height and velocity tracking, terrain traversal, mode transitions, and confined-space traversal. ### Source excerpt arXiv:2610.36546v1 Announce Type: cross Abstract: Humanoid locomotion in confined spaces requires continuous height adaptation across multiple locomotion modes, yet most existing approaches remain limited to height-adjustable bipedal walking. We present REACH, an RL-based hierarchical framework that enables a Unitree G1 humanoid to traverse height-constrained environments through variable-height bipedal walking and crawling. The learned policies regulate body height over complementary command ranges of 1.27--0.50m in bipedal locomotion and 0.50--0.20m in crawling, extending traversal beyond the clearance limits of crouched walking. To address contact-rich crawling dynamics, we employ a history-conditioned latent representation that captures terrain and interaction context from proprioceptive observations. A bidirectional transition policy further connects crouched bipedal locomotion and crawling across large kinematic and contact discontinuities. These skills are integrated with a LiDAR-based planner that converts point clouds into collision-free three-dimensional paths and locomotion commands, including proactive height targets. Real-world experiments demonstrate accurate height and velocity tracking, robust terrain traversal, reliable mode transitions and confined-space traversal with the proposed pipeline. Code and videos are available at https://aislab-sustech.github.io/REACH/. ## eMudhra quits global TLS certificate market after compromise of CA signing keys DevFeed: [eMudhra quits global TLS certificate market after compromise of CA signing keys]() Original publisher: [Read original article]() Author: Gyana Swain Published: 2026-10-23T10:27:24Z Content type: news Language: en Sources: [Computerworld]() Topics: [certificates](), [TLS (Transport Layer Security)](), [Certificate Transparency](), [enterprise certificate authority certificates](), [trust](), [digital sovereignty]() Tags: [browser-security](), [browser-security-browsers-security](), [browsers](), [ca](), [certificate](), [certificate-authorities](), [certificates](), [compliance](), [compromise](), [cost](), [critical](), [customer](), [data-sovereignty](), [digital-trust](), [edition-us](), [emudhra](), [g1](), [gartner](), [global](), [incident](), [keys](), [language-en](), [organizations](), [private-keys](), [security](), [storytype-news](), [tls]() ### AI overview eMudhra is exiting the global SSL/TLS certificate business after private keys for eight intermediate CAs were compromised. It offers customers free Sectigo-issued replacement certificates and says it will maintain revocation and certificate-status services until its issued certificates expire or are replaced. CertiNext users can request replacements in their accounts; other customers can contact support. eMudhra says its G1 RSA roots remain active during the transition, but analysts describe them as a possible bridge and urge migration of critical services to an independent, trusted CA. Google said removal of the affected emSign root from Chrome's trust store was expected October 27; sites chaining to it must replace certificates to avoid breakage. Separate fixes may be needed in Java and other systems with their own trust stores: Azul updated its JDK builds to block the eight certificates, while Java does not check revocation by default. The excerpt characterizes residual risk as limited and targeted, and notes that some organizations may not know where certificate trust decisions occur. ### Source excerpt Indian digital trust company eMudhra has announced that it is exiting the global SSL/TLS certificate business and is offering customers certificates issued by Sectigo as an alternative, following the compromise of private keys behind eight of its intermediate certificate authorities. The company's board decided "to voluntarily withdraw from the global SSL/TLS certificate business," eMudhra said in a press note. It added that data sovereignty and compliance requirements had increased the business's complexity and cost. "The level of social media amplification around this incident has brought scrutiny that is out of proportion to SSL/TLS's place in our business," eMudhra Chairman V Srinivasan said in the announcement. The company said it will contact affected customers to replace certificates free of cost. It also said it will keep revocation and certificate status services running until every issued certificate has expired or been replaced. Sectigo offered as an alternative In the customer update, eMudhra said it is making Sectigo-issued certificates available to all its customers. "Our G1 root certificates remain active and continue to be fully supported," the company said in the update. The October 22 release said the G1 RSA roots "will continue to serve customers until this transition is complete." eMudhra said it would provide replacement certificates free of charge through its partners and continue operating revocation and certificate-status services until all certificates it issued had expired or been replaced. CertiNext users can request Sectigo certificates through their accounts; other customers can contact eMudhra support, the statement added. Niyati Daftary, principal analyst at Gartner, said organizations should view the G1 roots "as a possible bridge, not automatically as the final destination." "With the company exiting global SSL/TLS and its remaining roots still subject to browser scrutiny, the prudent course is to move critical services to an indepen