Locomotion Latest coverage

MUSBRIC: LEGO-Compatible Pneumatic Muscle Bricks for Reconfigurable Musculoskeletal Robots

This paper introduces MUSBRIC, a modular hardware platform featuring LEGO-compatible pneumatic muscle bricks designed for reconfigurable musculoskeletal robots. The system includes a MuJoCo-based simulator validated against experimental actuator behavior. Demonstrations include an elbow-joint prototype showing load-dependent motion and a quadruped walker assembled from four legs. The project aims to facilitate systematic exploration of attachment geometries and functional robot construction, with hardware designs and simulation code planned for open-source release.

Zhiyuan Advances Embodied AI from Demonstration to Deployment with G2 Max Robot

At the 2026 Zhiyuan Innovation Day forum in Shanghai, Zhiyuan presented its transition from stage demonstrations to industrial deployment. The company unveiled the GO-3 Preview foundation model, Genie Master agent framework, and the G2 Max heavy-load industrial robot. Specific metrics highlighted improvements in task success rates through collective experience learning and post-training scaling. Early deployments are occurring in electronics manufacturing and retail logistics settings.

Robotic Systems Lab@leggedrobotics· 10:12SignalEditorial score6565

The ETH Robotic Systems Lab introduces 'Anticipated Perception for Humanoid Locomotion,' a new method that prioritizes terrain perception for the feet's upcoming landing points, aiming to enhance stability and navigation in humanoids.

nvidia/Metatron-v1

NVIDIA has released ONNX inference exports for Metatron v1, a unified policy model for the Unitree G1 humanoid robot that handles walking, scene interaction, and recovery. The release includes two artifacts: a single-robot demo pipeline (demo/model.onnx) and a six-robot batched network for G1 Racer (racer/network.onnx). Both files use learned policy weights from epoch 6410 but have different runtime interfaces and are not interchangeable. The model supports deployment beyond the browser, including other simulators and real robots, though hardware use requires independent validation. NVIDIA emphasizes that these exports are not complete training packages, do not include datasets or checkpoints, and are subject to the NVIDIA Open Model License Agreement.

DiffContact: Differentiable Contact Dynamics for Policy Learning and Contact-Implicit Trajectory Optimization

This research presents DiffContact, a method providing smooth analytic gradients for frictional contact to facilitate policy learning and contact-implicit trajectory optimization in legged locomotion. By employing a central-path relaxation of the de Saxcé-corrected second-order-cone complementarity problem, the approach ensures continuous sensitivity through contact-mode transitions. The authors demonstrate that training with 100,000 samples on a desktop PC yields policies that transfer directly to hardware, where the robot successfully withstands pushes and traverses challenging terrains like debris and slippery ground. Additionally, the method achieves contact-rich solutions in trajectory optimization that were previously missed by exact or Signorini-relaxed gradients.

Dispersive Tree Agreements for Multi-Robot Coordination

This research introduces WarpWeaver, a decentralized, real-time planner that approaches joint optimality for multi-robot nonlinear systems by using dispersive trees of forward-propagated commands and negotiation driven by decentralized column generation. The method replaces inter-robot coupling with local constraints, allowing locally grown dispersive trees to contain near-optimal agreements. In simulation on the Dynobench suite, WarpWeaver reduces total travel time by up to 22% compared to centralized offline planners and by up to 16% against a decentralized real-time baseline, while outperforming task-specific planners on networked navigation. Hardware experiments on six robots planning onboard demonstrate real-world implementability.

Memory before Goal: Role of Pre-History Memory for Humanoid Loco-Manipulation

This paper presents WholeBodyMem, an object-guided perceptual memory system that retains pre-history information for humanoid loco-manipulation in dynamic environments. By using SAM3 to track task objects and causally selecting key frames within an adaptive budget, the action expert integrates these frames via cross-attention with object tokens encoding appearance, geometry, and time. The authors also release HuMeGo, a benchmark featuring six simulated and six real-world Unitree G1 tasks. WholeBodyMem improves success rates from 29.1% to 72.6% in simulation and from 36.7% to 88.3% on the real robot, while also enhancing performance on standard SIMPLE tasks from 70.2% to 86.9%.

Do Earth-Trained Humanoid Controllers Walk on the Moon? Gravity, Terrain, and Soil Transfer

This research investigates how humanoid robot locomotion controllers trained on Earth perform under lunar gravity, terrain, and soil conditions.

KAIST Humanoid v0.9 Robot Pushes 77 Pounds and Lifts 44 While Walking

KAIST's Humanoid v0.9 uses deep reinforcement learning and onboard vision to coordinate walking with heavy loads, pushing 35 kg (77.2 lb) and lifting 20 kg (44.1 lb) while walking without external cameras. The demonstration highlights the challenge of whole-body locomanipulation, balancing forces during physical interaction with objects.

Rollbot 2: Agile Locomotion by a Single-Actuator Spherical Robot

Rollbot 2 improves upon the original Rollbot by maintaining a single-actuator design while introducing a redesigned mass distribution for more direct movement. It uses onboard inertial and ultra-wideband sensing to autonomously navigate toward beacons and follow waypoints, achieving a 468-meter indoor route at 0.87 m/s and completing a 4.8-meter 'N' benchmark in 10.8 seconds with a mean deviation of 7 cm.

Push the Limit: Learning High-Speed Sprinting and Aggressive Cornering on a Full-Scale Humanoid

This paper presents a learning-based framework that enables a full-scale bipedal humanoid to sprint at over 7 m/s with motion-prior-guided gait and perform aggressive cornering at high speed with a pronounced trunk lean of up to 25°. The framework addresses challenges like high-frequency yaw-rate oscillations and skill-specific reward design through causal multi-head attention and skill-conditioned discriminators, achieving stable high-speed sprinting and aggressive cornering in both simulation and real-robot deployment.

Soft Robotic Fish Combines High-Speed Swimming with 3D Maneuverability

Researchers from the Shenyang Institute of Automation have developed a soft robotic fish that can swim at high speeds and perform agile three-dimensional maneuvers. The design integrates a rigid–flexible coupled propulsion system with dynamic buoyancy regulation, inspired by biological mechanisms such as muscle-skeleton coordination and swim bladder function. The robotic fish achieves a maximum speed of 2.13 body lengths per second and can perform complex 3D movements like turning, ascending, and spiral ascent, offering a potential approach for high-performance underwater robots.

Next-gen hopping robot is the smart flea we didn't know we needed

Engineers at the University of Washington have developed DirectHop, a hopping robot inspired by the flea's energy-efficient jumping mechanism. The robot can jump up to 50 times its body height and is controlled via an external power source. Future versions may include onboard batteries and sensors, with the goal of creating an independent robot-explorer for tasks like environmental monitoring and inspection.

Feather star-inspired underwater robot gets 3D maneuverability from only two actuators

Researchers have developed an aquatic soft robot inspired by feather stars that can move in all three dimensions using only two actuators. The design utilizes mechanical intelligence to achieve efficient 3D maneuverability through three modes: jellyfish, fish, and rotor. The robot can explore underwater spaces, lift objects, and work collaboratively with other robots, with future plans to develop a fully wireless version.

Robotic Systems Lab@leggedrobotics· 16:29SignalEditorial score7878

The team at ETH Robotic Systems Lab demonstrated high-speed rough-terrain locomotion on ANYmal using an automatic curriculum learning approach called LP-ACRL. This method dynamically samples terrain types, difficulty levels, and velocity commands based on policy performance, without requiring pre-defined training sequences. The work was presented at RAL 2026.

Small robot propulsion: low-power twisting enables hopping, swimming

Researchers at the University of Michigan and UCLA have developed a new propulsion mechanism for small robots using low-power twisting of elastic rods. This method enables hopping and swimming motions by utilizing snapping mechanics. The mechanism works on challenging surfaces such as slippery cloth and grass-covered ground, producing frog-like motion. The study, published in Science Advances, was co-led by Khalid Jawed and Xiaonan (Sean) Huang, and demonstrates the potential for miniature robots to maneuver over uneven or slippery terrain.

F.03 Arrives at BMW

Figure's latest generation robot, F.03, has arrived at BMW Group Plant Spartanburg, demonstrating advanced dynamic whole-body control and sequencing use cases. Powered by Helix 02, the robot performs complex material manipulation tasks that require both precision and force, marking progress in embodied AI for manufacturing.