# Vision-Language-Action Models

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

Vision-language-action policies process hundreds of visual tokens per step, which drives up inference latency. APS-VLA keeps the backbone frozen and uses a lightweight scorer conditioned on the current observation, instruction and proprioception to pick tokens, training the retained representation both to reproduce policy outputs and to predict the subsequent visual change. On LIBERO with 128 tokens retained, mean success is 95.75%, 1.25 points below the unpruned 97.00%, and inference latency falls to 68.75 ms.

In humanoid manipulation, a large vision-language-action backbone is slow, so synchronous execution acts on stale observations and control frequency falls as the backbone grows. DASH splits the policy into a slow vision-language model and a fast flow-matching action head running as three concurrent processes over LCM, so the control rate is set by the action head alone. On a RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz from 0.8B to 8B backbones while the synchronous rate falls from 9.7 Hz to 4.1 Hz, with higher success at every size; on a real dexterous humanoid, pick-and-place succeeds in 23 of 24 trials versus 16 of 24 for the synchronous pipeline.

In humanoid manipulation, a large vision-language-action backbone is slow, so synchronous execution acts on stale observations and control frequency falls as the backbone grows. DASH splits the policy into a slow vision-language model and a fast flow-matching action head running as three concurrent processes over LCM, so the control rate is set by the action head alone. On a RoboTwin benchmark that charges inference latency in simulated time, DASH holds 20 Hz from 0.8B to 8B backbones while the synchronous rate falls from 9.7 Hz to 4.1 Hz, with higher success at every size; on a real dexterous humanoid, pick-and-place succeeds in 23 of 24 trials versus 16 of 24 for the synchronous pipeline.

DexForesight injects predictive features from an action-conditioned spatial world model into a dexterous policy's latent space through query alignment, and at deployment keeps only the policy, which acts from RGB, language and robot state. On 11 DexJoCo simulation tasks it raises the GR00T N1.7 backbone's overall success from 49.4% to 60.5%, and on three real-robot tasks it reaches 58.3%, 8.3 points above the strongest baseline. The world model, depth encoder and future-prediction heads are all removed at inference.

DexForesight injects predictive features from an action-conditioned spatial world model into a dexterous policy's latent space through query alignment, and at deployment keeps only the policy, which acts from RGB, language and robot state. On 11 DexJoCo simulation tasks it raises the GR00T N1.7 backbone's overall success from 49.4% to 60.5%, and on three real-robot tasks it reaches 58.3%, 8.3 points above the strongest baseline. The world model, depth encoder and future-prediction heads are all removed at inference.

When a robot draws on past experience, matching usually relies on task or scene similarity, which cannot separate execution stages such as approaching a bowl from having already placed something in it. Robota-Embed encodes language, vision and native action together in both the current-state query and the recorded future clip, so motion enters the retrieval key itself. On Robota-Embed-Bench, 4,000 queries against a shared 8,000-clip gallery, it reaches 47.33% clip-level and 81.77% episode-level R@1; the abstract reports gains of 22.09 and 33.47 percentage points over the action-free Gemini Embedding 2, while Section 4.1 gives 22.09 and 33.18 points.

When a robot draws on past experience, matching usually relies on task or scene similarity, which cannot separate execution stages such as approaching a bowl from having already placed something in it. Robota-Embed encodes language, vision and native action together in both the current-state query and the recorded future clip, so motion enters the retrieval key itself. On Robota-Embed-Bench, 4,000 queries against a shared 8,000-clip gallery, it reaches 47.33% clip-level and 81.77% episode-level R@1; the abstract reports gains of 22.09 and 33.47 percentage points over the action-free Gemini Embedding 2, while Section 4.1 gives 22.09 and 33.18 points.

Ai2 released ToyBlocks, an open simulation framework built around a robot arm and tabletop blocks, the first controlled world in its FoundationWorlds program. It automatically generates tasks and demonstrations, letting researchers vary training data, repeat experiments and trace why a robot fails. In physical tests, the low-level skill model ManiFlow, trained only on simulation data, reached 80.0% success in simulation and 81.1% on the real robot, while the combined high-level planner plus low-level executor succeeded in 55.5% of simulated episodes and 50.0% of real-world episodes.

Ai2 released ToyBlocks, an open simulation framework built around a robot arm and tabletop blocks, the first controlled world in its FoundationWorlds program. It automatically generates tasks and demonstrations, letting researchers vary training data, repeat experiments and trace why a robot fails. In physical tests, the low-level skill model ManiFlow, trained only on simulation data, reached 80.0% success in simulation and 81.1% on the real robot, while the combined high-level planner plus low-level executor succeeded in 55.5% of simulated episodes and 50.0% of real-world episodes.

The highest-scoring action from a vision-language-action (VLA) model is not always the most reliable one to execute. FRS-Grounding keeps the finite candidate set after VLA inference, fuses normalized candidate scores with distance to a safe-action set into a fuzzy action set, and builds a pairwise relation from pose differences and collision-free states before fuzzy rough lower and upper approximations separate strongly supported candidates from boundary cases. Across 1500 randomized simulations the full pipeline reached 86.4% success and a 2.1% safety-violation rate, versus 74.8% and 8.2% for the pointwise hard-constraint baseline VLA-HC; under positional noise with σ = 10 mm it retained about 78% success against about 28% for VLA-HC.

The highest-scoring action from a vision-language-action (VLA) model is not always the most reliable one to execute. FRS-Grounding keeps the finite candidate set after VLA inference, fuses normalized candidate scores with distance to a safe-action set into a fuzzy action set, and builds a pairwise relation from pose differences and collision-free states before fuzzy rough lower and upper approximations separate strongly supported candidates from boundary cases. Across 1500 randomized simulations the full pipeline reached 86.4% success and a 2.1% safety-violation rate, versus 74.8% and 8.2% for the pointwise hard-constraint baseline VLA-HC; under positional noise with σ = 10 mm it retained about 78% success against about 28% for VLA-HC.