News without the noise
42 results for “LIBERO” · page 2 of 3
Research · Robotics & embodied AI
Native Action-Prior Learning from Videos for World Action Models
NAVA-WAM uses visual-transition prediction to train an action model before adapting it with action-labeled robot demonstrations.
LIBERO: 99% success rate (as reported)
Meta AI, University of Copenhagen, Physical Intelligence
Research · Robotics & embodied AI
GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies
GeoAAC uses the geometry of flow matching denoising trajectories to adjust the action chunk size per step, boosting real-world success rates from 53.3% to 74.4%.
real-world manipulation tasks: 74.4% success rate (as reported)
Tongji University, The Hong Kong Polytechnic University, University of Electronic Science and Technology of China
Research · Robotics & embodied AI
StructRL: Online Structured Reinforcement Learning for Long-Horizon Vision-Language-Action Tasks
StructRL rewards verified subtasks in prerequisite order, giving robot policies useful feedback before an entire manipulation sequence succeeds.
LIBERO Goal: 95.6% success rate (as reported)
The Pennsylvania State University, Amazon AWS AICode
Research · Robotics & embodied AI
InternW0-$Δ$: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
InternW0-Delta combines video, semantic, geometric and action models, trained on more than 20,000 hours of heterogeneous demonstrations.
LIBERO-Plus: 92.8% success rate (as reported)
Shanghai AI Laboratory
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: 95.9% 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: 0.83 state-probe relative position error (as reported)
University of California, Santa Cruz, University of California, Santa Barbara, Google
Research digest
The week’s clearest result is that reliable agents require runtime constraints, process-aware evaluation, and targeted correction—not stronger instructions alone.
Research digest
The strongest results move reliability out of prompts and into authorization layers, persistent state, structured interfaces, and explicit credit assignment.
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): 0.09 pts NBT (as reported)
University of Electronic Science and Technology of China, Shanghai AI Lab, Fudan University
Research · Robotics & embodied AI
Rolling-WAM: World Action Models with Rolling Imagination
The method reuses partially denoised video and action predictions across control cycles instead of regenerating the full planning horizon.
LIBERO: 98.1% success rate (as reported)
University of Southern California, Brown University, Fudan University
Research · Robotics & embodied AI
VLA-HWBench: Where the Time Goes When VLA Models Run on Edge Hardware
Benchmarking 18 models with a seven-stage inference breakdown reveals that kernel fusion and few-step generation are key to real-time robot control.
LIBERO-10: 318 ms latency (as reported)
National University of Singapore, 0G Labs
Research · Robotics & embodied AI
VLA-ZO: Fast Zeroth-Order Adaptation for Vision-Language-Action Models
VLA-ZO accelerates memory-efficient adaptation by updating only action components and reusing vision-language conditioning across parameter perturbations.
LIBERO camera-viewpoint shifts: 58.17% average task success (as reported)
UNIST
Research · Robotics & embodied AI
ReCAT: Remember, Count, and Time: Structured Recurrent Memory for Robot Manipulation
ReCAT combines structured recurrent memory with direct current observation to achieve high success on benchmark and real-world tasks
LIBERO: 95.3% success rate (as reported)
University of Edinburgh, Karlsruhe Institute of Technology, NVIDIA
Research · Robotics & embodied AI
Sensorimotor Alignment: A Platonic Proxy for Embodied Policy Success
A training-free metric measures how well sensory representations match motion trajectories, correlating with closed-loop performance across model families and settings.
Meta-World, Robosuite, Franka-Kitchen: 0.77 Pearson r (as reported)
Rice University, NVIDIA
Research digest
Across agents, training and evaluation, the strongest evidence separates improving a proxy from improving the system that practitioners actually deploy.
Research · Robotics & embodied AI
Remembering What Matters for Long-Horizon Robotic World Models
ReCAP method preserves initial scene and recent frames while discarding stale history, boosting fidelity and efficiency without retraining.
LIBERO-90: 12% LPIPS curve AUC reduction (as reported)
The Hong Kong University of Science and Technology (Guangzhou), North China Electric Power University, LIGHTSPEED
Research · Robotics & embodied AI
Robota-Embed: Unified VLA Embeddings for Robot In-Context Learning and Evolution
Robota-Embed matches a robot’s current task and execution stage to useful recorded continuations, improving retrieval and downstream policy performance.
LIBERO difficult subset: 92.8% success (as reported)
NVIDIA, MIT, UCB
Research · Robotics & embodied AI
BIND: Binding 3D Robot Actions to 2D Image Features
BIND links candidate 3D end-effector positions to projected image features, avoiding the need to learn this spatial correspondence from demonstrations.
LIBERO: 92% success rate (as reported)
University of Southern California, Toyota Research Institute