News without the noise
Benchmark
Reported in 31 papers on Zotpaper. The latest, each as the paper states it:
LIBERO-Plus (16)LIBERO-PRO (6)LIBERO-Long (3)All benchmarks →
26 results for “LIBERO” · page 1 of 2
Research · Robotics & embodied AI
Video Prediction Policy 2: Predict Better, Act Better
VPP2 combines manipulation-focused video pretraining with action learning while preserving the video model’s ability to generalize.
LIBERO-ID: 98.8% overall success rate (as reported)
Robotera, Tsinghua University, HKUST (GZ)
Research · Robotics & embodied AI
Embodiment-Classifier Guidance: Transferring Frozen VLAs Across Robot Dynamics without Demonstrations
Han and Niranjani steer a frozen flow-based VLA with a lightweight embodiment classifier, raising success from 72% to 80% on modified LIBERO controllers.
LIBERO: 72% success rate (as reported)
UC Berkeley
Research · Robotics & embodied AI
Trace-WAM: Learning to Act via 3D Motion Prediction for World Action Models
The method, Trace-WAM, jointly predicts future video, 3D motion, and actions, improving performance on several benchmarks and real-world tasks, especially when data is limited.
LIBERO: 98.7% success rate (as reported)
The Hong Kong Polytechnic University, Alibaba Group, Tsinghua University
Research · Robotics & embodied AI
PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
PDRVLA extracts compact physical tokens from VLM features via future prediction objectives, achieving high success rates on four simulation benchmarks without online video generation.
LIBERO: 99.3% success rate (as reported)
Peking University, JD.com
Research · Safety & security
Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution
R2A retrieves motion primitives from past demonstrations and fuses them with frozen policy actions, lifting success rates by up to 12 points in simulation and 40 points on a real robot.
LIBERO-Plus: 98.6% success rate (as reported)
VinUniversity, VinRobotics, Hanoi University of Science and Technology
Research · Robotics & embodied AI
Leap-VLA: High-Rate Robot Control from a Low-Rate VLM Backbone
Cached semantic context plus a fast visual stream allows robot control at 20-50 Hz with 1-second VLM updates, recovering over 94% success rate.
LIBERO Goal: 97.2% success rate (as reported)
Leapmotor
Research · Robotics & embodied AI
REKI: Registering Knowledge in VLAs
REKI combines public-data pretraining with cached register tokens that preserve recent visual context at less than 1 percent inference overhead.
LIBERO: 21.8 ms inference latency per action chunk (as reported)
Kyutai, Inria, École nationale des ponts et chaussées
Research · Robotics & embodied AI
RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions
Zhao and colleagues pretrained a robot policy on motion representations extracted from 30,100 hours of video, then adapted it to specific robots and evaluated manipulation performance.
LIBERO: 98.8% success (as reported)
Hong Kong Embodied AI Lab, CUHK, PUMC
Research · Robotics & embodied AI
RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions
Zhao and colleagues report that a model trained across human, robot and simulation videos generates additional demonstrations that improve robot policy success in the evaluated settings.
LIBERO: 96.1% success rate (as reported)
Hong Kong Embodied AI Lab, CUHK, DAMO Academy, Alibaba Group
Research · Efficiency & systems
Serving Agentic Robot Fleets: Workload Characterization, KV-Cache Thrashing, and Admission Control
In sustained replay of robot workloads, limiting active episodes restored shared computing throughput and roughly quadrupled completed programs.
LIBERO-PRO: 7.78 s Astra request latency (median) (as reported)
Huawei Inc.
Research · Robotics & embodied AI
TacitVLA: Reasoning Vision-Language-Action Models Without Reasoning Latency
TacitVLA matched explicit-reasoning policies across four robotics and driving benchmarks while reducing action latency by up to 3.4 times.
LIBERO: 98.6% Success (as reported)
NVIDIA, New York University, University of California, Berkeley
Research · Robotics & embodied AI
Long-WAM: Scaling the Context of World-Action Models
Long-WAM scales visual history up to 19.2 seconds for action prediction, achieving high success on dynamic tasks.
LIBERO: 99.5% average success rate (as reported)
NVIDIA, MIT, HKU
Research · Robotics & embodied AI
Wiring Matters: Injection Topology and Initialization of Affordance Heads in Vision-Language-Action Policies
A stop-gradient and an actively initialized residual bridge lift LIBERO success from 93.1% to 96.2%, matching far more complex architectures.
LIBERO: 91.4% long-horizon success rate (as reported)
Drexel University, Virginia Tech, Amazon Store Foundation AI
Research · Generative media
RealtimeWAM: One-Step Asynchronous World Action Models
RealtimeWAM uses teacher-anchored consistency distillation and cross-expert pipelining to accelerate Mixture-of-Transformers based world action models by up to 25x on H100 GPUs with less than 1% accuracy loss
LIBERO: 99% success rate (as reported)
Nanyang Technological University, Beihang University, Sensetime
Research · Robotics & embodied AI
SUAVE: Unified Video-Action Models via Masked Diffusion
SUAVE uses masked diffusion over shared discrete tokens to generate video, robot actions, or both with the same model weights.
LIBERO: 95.4% success rate (as reported)
Columbia University, Toyota Research Institute
Research · Robotics & embodied AI
Seeing Through the Illusion: Test-Time Scaling for World Action Models
A frozen world model that generates both actions and predictions can be made four times more reliable by evaluating candidates on action validity, inverse-dynamics consistency, and task progress before acting.
LIBERO: 3.5 x relative episode runtime (as reported)
Huawei Technologies Canada, University of British Columbia
Research · Robotics & embodied AI
TAPESTRY: A Single Backbone for Video Generation and Robot Control
The authors show that representing robot actions as single-patch tokens in a pretrained video diffusion transformer's latent space allows real-time action generation without video synthesis.
LIBERO: 70% average success rate (as reported)
Research · Robotics & embodied AI
W2A: Distilling World Priors into Action Models for Real-Time Robotic Control
A two-stage training method transfers what a video prediction model has learned about how scenes unfold into a robot control model, then throws the video model away before deployment.
LIBERO: 93.6% success rate (as reported)
Harbin Institute of Technology, Shenzhen
Research · Vision
ALPS: Visual State Encoders for World Models
ALPS, a ViT-B/16 pretrained from still images and text, matches the state-probe error of a 3.5x larger model while preserving language alignment and pixel decode ability
LIBERO: 85.5% mean policy success (as reported)
University of California, Santa Cruz, University of California, Santa Barbara, Google
Research · Robotics & embodied AI
Data-Free Continual Learning for World Action Model with Rectified Generative Replay
Fang et al. show that rectified generative replay, which truncates rollouts at task completion and filters physically inconsistent ones, reduces forgetting by over 90% versus sequential fine-tuning.
LIBERO (average): 94.6% SR (as reported)
University of Electronic Science and Technology of China, Shanghai AI Lab, Fudan University