DeepMind introduces Gemini Robotics 2.5 for robot dexterity
DeepMind reports 30% to 100% higher success, faster tasks, and 96% kitting success on Mercedes-Benz lines.
Stories per day · 30 days
1 RSS
Relatedrobotics24reinforcement-learning6world-models6embodied-ai5libero5benchmark3manipulation3robotwin3vla3behavior-cloning2humanoid-robots2imitation-learning2robot-learning2simulation2affordance1alpamayo1apptronik1asimov-jailbreak1
48 stories, newest first · one card per story
Stories tagged #vision-language-action
A2-MEAS embodied framework reports BEHAVIOR and real-robot scores
Authors report 0.6411 Q-score on 100 public tasks and 57.92% success on 12 real-robot tasks versus 26.67% for pi0.5.
NEO-VLA: Native Vision-Language-Action Framework Outperforms Encoders
Native vision-language-action models achieve stronger robotic control by evolving visual representations jointly with language.
PolicyIK provides high-precision inverse kinematics via reinforcement learning
PolicyIK uses goal-conditioned RL to achieve millimeter-precision IK for robot arms, outperforming classical controllers and enabling 95.7% success in simulated pick-and-place with π0.5.
Gym-Anything builds RoboWorld, 25 coding-agent robotics simulations
Frontier models score an average reward of 0.06 and fail every RoboWorld test task; several teaching methods raise performance.
Q-adaptive guidance improves reasoning vision-language-action driving
Frozen Alpamayo 1.5 with the method scores 8.135159 on the hidden Waymo set, above 8.079.
OpenFlyV2 flies a quadrotor onboard from vision and language
Authors report 48% zero-shot navigation success in four real environments with no real aerial training data.
HiFOR hierarchical foresight with action-guided world grounding
HiFOR jointly predicts short-term states semantic milestones and goals using hierarchical dependency mask and action-guided retrieval for vision-language-action models.
Underwater VLA policies beat baseline; ACT beats SmolVLA at sea
A 14-model DynamicVLA study and an ROV case study at a Malta wreck also open-source the BLUEMBODIED KIT.
ThinkWorld trains robot action reasoning from unlabeled human video
Chain-of-Dynamics tokens let human and robot video supervise the same action-reasoning space.
Just Action Models match world-action policies without visual generation
Against FastWAM, reported LIBERO success moves from 97.0%, 55.3%, and 91.2% to 98.2%, 64.5%, and 91.4%.
PDRVLA conditions robot actions on physical dynamics without video generation
It reports 99.3% success on LIBERO and 81.5% on SimplerEnv using public data and a frozen world model.
SO-101 Bench benchmarks language-conditioned spatial robot manipulation
A simulated twin matched the real arm's policy ranking exactly, though simulation overestimated success.
RobustRLlib benchmarks 17 robust reinforcement learning algorithms
Evaluation across 273 perturbation settings maps where current robust RL methods still fail.
Backtrack lets frozen vision-language-action policies retry failed motions
On LIBERO-Long, failure rates fall 19–48% and success rises by up to 7.4 points; the head adds 0.48–0.88% parameters.
DiscoDemo generates diverse robot demos to improve imitation learning
Authors report downstream success rising from 40% to 62% in simulation and from 39% to 78% on a real robot.
Echo denoises stop and failure signals inside vision-language action policies
Echo learns task status without frame-level progress labels and uses those signals to condition later action denoising.
FocusVLA routes multi-view inputs using predicted future intent
On RoboTwin 2.0 it raises average success 10.10 points over π0.5, and real-world success from 71.7% to 85.0%.
HiSTEVE 1.0 uses a hierarchical VLA agent for open-world Minecraft
It scores 85.0% on MineStudio Bench and 37.4% average success on OpenHA, and can craft an iron pickaxe from scratch.
SP-HLM links sparse poses to whole-body humanoid control
A vision-language-action model predicts five sparse poses, then a tracker combines them with a learned motion prior.
AW-0 couples code planning with video and action prediction
The paper reports better RoboTwin 2.0 and bimanual task results than VLA and WAM baselines.
UniWALM unifies video, action and language for robot control
Authors report 34.5% on RMBench memory tasks without a pretrained language model, and Find-the-Item success rising from 11.1% to 61.1%.
VLDA learns compact deltas as robotic action interfaces
Implemented in LoongForge, it targets long-horizon consistency without predicting full future states.
Meet Kyutai REKI: The 2.4B VLA Model Bringing Efficient Memory to Robot Manipulation
Kyutai introduced REKI, a 2.4B vision-language-action model for memory-efficient robot manipulation.