Releases·Model Release
Unitree releases UnifoLM-WLA-1.0: a 6B humanoid VLA trained on 2,500 hours of real robot data, running 64 tasks on the G1
Unitree Robotics released UnifoLM-WLA-1.0, a 6B-parameter general-purpose humanoid foundation model trained on ~2,500 hours of real robot data. A single model handles 64 tasks (10 whole-body, 54 tabletop) on the Unitree G1, including laundry and bed-making, and supports parallel grippers plus two dexterous hands. It is one of the more complete open attempts at a whole-body VLA to date.
7 engagement1 sourceSat, Oct 3, 2026, 12:00 AM
AI Overview
- Unitree's UnifoLM-WLA-1.0 is a 6B-parameter general-purpose humanoid foundation model trained on ~2,500 hours of real robot data
- One model handles 64 tasks (10 whole-body, 54 tabletop) and supports parallel grippers plus two dexterous hands, per the project page (unigen-x.github.io)
- The model starts from UnifoLM-ER-1 (a Qwen3-VL-based embodied reasoner), adds optical-flow future dynamic region prediction with VQ-VAE, residual-VQ action discretization, and an MMDiT action expert for continuous control
- Demos on the Unitree G1 include making the bed, loading the washing machine, folding clothes, and sorting objects; the team reports strong spatial reasoning on embodied benchmarks vs. open-source models
Related
Developments·Benchmark
HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution
Researchers introduce HumanoidToolBench, an 18-task benchmark evaluating humanoid tool use across tool selection, manipulation, and locomotion, plus ToolBook, a 3.1k-demonstration dataset collected in simulation and on a real Unitree G1. Evaluations of seven simulated and three real-robot policies reveal large gaps between selecting a suitable tool and actually completing tasks. The benchmark gives robotics practitioners a standardized way to measure whole-body tool-use capabilities that existing benchmarks cover only in isolation.
Researchers introduce HumanoidToolBench, an 18-task benchmark evaluating humanoid tool use across tool selection, manipulation, and locomotion, plus ToolBook, a 3.1k-demonstration dataset collected in simulation and on a real Unitree G1. Evaluations of seven simulated and three real-robot policies reveal large gaps between selecting a suitable tool and actually completing tasks. The benchmark gives robotics practitioners a standardized way to measure whole-body tool-use capabilities that existing benchmarks cover only in isolation.
0 engagement1 sourcearxivThu, Oct 1, 2026, 05:21 PM
Developments·Paper
TANGO: Whole-Body Vision-Language Model Navigates Unitree G1 Humanoid Through Cluttered Environments Zero-Shot
A research paper introduces TANGO, the first whole-body vision-language navigation framework for language-conditioned humanoid traversal of cluttered indoor environments. It treats navigation as continuous geometry-aware whole-body adaptation rather than 2D path planning, coordinating arms, torso, and gait for collision-free movement in complex 3D spaces.
A research paper introduces TANGO, the first whole-body vision-language navigation framework for language-conditioned humanoid traversal of cluttered indoor environments. It treats navigation as continuous geometry-aware whole-body adaptation rather than 2D path planning, coordinating arms, torso, and gait for collision-free movement in complex 3D spaces.
0 engagement1 sourcearxivTue, Sep 8, 2026, 05:59 PM
Releases·Model Release
Xiaomi releases XR-1 robotics foundation model trained on 100K hours
Xiaomi released Xiaomi-Robotics-1 (XR-1), a Vision-Language-Action (VLA) robot foundation model trained on over 100K hours of real-world manipulation trajectories. It uses a two-stage training paradigm—pre-training for breadth and post-training for alignment—to enable out-of-the-box mobile manipulation in unseen environments and efficient adaptation to new tasks.
Xiaomi released Xiaomi-Robotics-1 (XR-1), a Vision-Language-Action (VLA) robot foundation model trained on over 100K hours of real-world manipulation trajectories. It uses a two-stage training paradigm—pre-training for breadth and post-training for alignment—to enable out-of-the-box mobile manipulation in unseen environments and efficient adaptation to new tasks.
7 engagement1 sourceredditWed, Aug 5, 2026, 08:27 PM
Developments·Paper
Spectral skills: predictive latent motion representation cuts humanoid tracking error 62% on Unitree G1
A new paper introduces spectral skills, a latent command interface between planner and controller for humanoid robots, learned by predicting subsequent motion rather than reconstructing inputs. On a 29-DoF humanoid, conditioning on spectral skills cuts global tracking error by 62% versus the state of the art, and the frozen controller supports skill chaining and composition of unseen behaviors, demonstrated on Unitree G1 hardware.
A new paper introduces spectral skills, a latent command interface between planner and controller for humanoid robots, learned by predicting subsequent motion rather than reconstructing inputs. On a 29-DoF humanoid, conditioning on spectral skills cuts global tracking error by 62% versus the state of the art, and the frozen controller supports skill chaining and composition of unseen behaviors, demonstrated on Unitree G1 hardware.
0 engagement1 sourcearxivTue, Sep 29, 2026, 02:27 PM
Developments·Paper
Xiaomi unveils 38B-parameter U0 model for unified embodied synthesis
Xiaomi Robotics introduces U0, a 38-billion-parameter multimodal autoregressive model that treats embodied generation as an extension of foundation image and video generation. It aims to overcome limitations of existing methods that sacrifice pre-trained visual knowledge when adapting to robot data.
Xiaomi Robotics introduces U0, a 38-billion-parameter multimodal autoregressive model that treats embodied generation as an extension of foundation image and video generation. It aims to overcome limitations of existing methods that sacrifice pre-trained visual knowledge when adapting to robot data.
0 engagement1 sourcearxivMon, Jul 13, 2026, 02:57 PM