"Open Source"
Vision-Language-Action (VLA) models represent the frontier of general-purpose embodied intelligence, translating visual sensor observations and high-level language instructions into low-level joint and gripper commands for robotic manipulators. However, existing open-source VLA systems (such as OpenVLA and Octo) suffer from brittle execution pipelines: their training recipes often fail under visual domain shifts (lighting, novel background clutter, camera jitter), and their tokenized action heads lack explicit predictive world dynamics, leading to catastrophic failure when complex multi-step manipulation sequences hit unexpected physical contact.
On August 31, 2026, robotics researchers unveiled VLANeXt, a modular, research-oriented open-source VLA codebase engineered to establish reproducible, state-of-the-art baselines across the standard LIBERO manipulation benchmark suite. By introducing lightweight backbone options down to 0.8B parameters and pioneering a World-Action-Model (WAM) variant that couples future video generation with continuous action prediction, VLANeXt establishes new performance ceilings—achieving up to a 98.2% success rate on LIBERO and lifting robustness under severe physical perturbations by more than 14%.
Key Breakthroughs.
1. VLANeXt-S
: High-Performance Compact VLA at 0.8B Scale Most current VLA architectures require multi-gigabyte backbones (7B+ parameters), making real-time control loops (typically 20 Hz to 50 Hz) latency-prohibitive on edge robot computers:
Qwen3.5-0.8B Core Backbone: VLANeXt-S packages the modern architectural improvements of Qwen3.5 into an ultra-compact 0.8B parameter model. 96.7% LIBERO Success Rate: Despite operating at a fraction of the computational footprint of 7B baselines, VLANeXt-S retains near-parity with massive models, making high-speed on-device robot policy execution practical on single consumer GPUs.
2. VLANeXt-WAM
: World-Action Dynamics via Video Generation A fundamental limitation of classical action-tokenizing transformers is "blind execution"—predicting action vectors without imagining the physical consequences of those actions on the scene:
Generative Dynamics Coupling: VLANeXt-WAM integrates the WAN2.2-5B video generation backbone to simultaneously predict future visual scene evolution alongside continuous 7-DoF end-effector actions. Peak 98.2% LIBERO Benchmark: By grounding action selection in explicit physical simulation of upcoming frames, VLANeXt-WAM drastically reduces compounding trajectory errors, securing an industry-leading 98.2% task success rate.
3. Superior
Generalization Under LIBERO-plus Perturbations Real-world robotics environments deviate constantly from pristine laboratory simulation:
Robustness Under Seven Perturbation Classes: Evaluated on LIBERO-plus—which tests robot policies against camera pose noise, light fluctuations, novel textures, table heights, and object distractors—VLANeXt increased the average success rate from 69.6% (OpenVLA-OFT) to 83.9%. Direct Continuous Action Diffusion Heads: Replaces discrete action token discretization with continuous action diffusion heads, eliminating quantization discretization artifacts in delicate insertion and grasping tasks.
Technical Specifications & Benchmark Overview
& Benchmark Overview Model Variant Backbone Scale Base Architecture LIBERO Avg Success LIBERO-plus Robustness Control Latency OpenVLA-OFT 7B Llama-2 / SigLIP 92.1% 69.6% ~85 ms Octo Baseline 100M Custom Transformer 78.4% 52.1% ~25 ms VLANeXt-S 0.8B Qwen3.5-0.8B 96.7% 79.4% ~22 ms VLANeXt (Standard) 7B Qwen / SigLIP2 97.4% 83.9% ~60 ms VLANeXt-WAM 5B WAN2.2 Video Diffusion 98.2% 86.1% ~75 ms
Verified Integration & API Usage
Developers can deploy VLANeXt policies into robotic control loops via Python:
pythonimport torch from vlanext import VLANeXtPolicy, RobotObservation # Initialize the lightweight 0.8B VLANeXt policy for on-robot deployment policy = VLANeXtPolicy.from_pretrained("cavanloy/vlanext-s-libero", device="cuda", dtype=torch.bfloat16) # Provide current camera frame and language instruction observation = RobotObservation( rgb_image=camera_feed.get_latest_frame(), # [3, 224, 224] joint_states=robot.get_current_joint_positions(), instruction="Pick up the yellow bowl and place it on the top wooden shelf" ) # Predict next 7-DoF continuous action chunk (x, y, z, roll, pitch, yaw, gripper) with torch.inference_mode(): action_chunk = policy.predict_action(observation, horizon=8) # Execute continuous trajectory on robot hardware robot.step(action_chunk) print("Dispatched action trajectory chunk:", action_chunk.shape)