Robot VLA Research Centers on Execution Checks, Memory, and Injury Safety

One paper targets a data bottleneck in generalist VLA training: action labels are scarce for robots, while human egocentric manipulation video is abundant. The method learns masked latent action tokens with a disentangled VQ-VAE, using physical masks to separate foreground motion from scene background. A Prismatic-7B vision-language model then predicts those tokens before robot adaptation.

The reported gains are concrete. On LIBERO, the full method reaches 91.8% average success, ahead of OpenVLA at 76.5% and Diffusion Policy at 72.4%. On RoboTwin 2.0 dual-arm simulation, it reaches 67.7% average success across 10 tasks. The downstream setting uses about 50 trajectories per task, so the claim is about cheaper adaptation after unlabeled video pretraining.

Two papers treat VLA execution as a live control problem. DREAM-Chunk samples multiple candidate action chunks at test time, predicts their latent futures with a small world model, and executes the chunk whose predicted state best matches the observed rollout. On a precise insertion task under perturbation, it reports π0.5 success of 65%, compared with 10% for open-loop execution.

Object-centric residual reinforcement learning adds a different correction path. A frozen VLA supplies base actions, while a pose-based residual policy trained in simulation adjusts those actions on a real FR3 robot. Across five tabletop tasks, average real success is 76% with the residual policy and 42% for the base VLA. The residual actor is small: about 0.06 ms per GPU forward pass, while the VLA takes roughly 140 ms.

World models appear in both capability and security roles. Mem-World adds a wrist-view-centered surfel memory so a robot video model can retrieve useful past views when the wrist camera is occluded or moving quickly. On 34 DROID memory-stress replay trajectories, it improves third-view PSNR to 25.30, compared with 23.17 for Ctrl-World. Its simulated success estimates also correlate more closely with real-world success, with r=0.97 across five tasks.

SWAAP examines the same class of learned dynamics as an attack surface. It poisons selected next-state targets in a fine-tuning buffer while leaving states, actions, and rewards unchanged. The paper reports evaluation on TD-MPC2 and DINO-WM across DMControl, MyoSuite, and MetaWorld, and tests stealth against residual, CUSUM, and TRIM-style defenses. The available summary does not give numeric return drops, so the grounded takeaway is the attack construction and the evaluated threat setting.

RoboShackles focuses on refusal behavior before a robot acts. It builds 10,000 synthetic hazardous robot video clips from real DROID observations, covering hand and human direct harm plus fire, electrical, water, and falling risks. The test set has 1,200 samples, with 200 samples per category.

The evaluation rule is strict: an embodied foundation model (EFM) is safe only if it refuses the instruction or produces no executable action. Six tested EFMs, including Cosmos-Policy, DreamZero, LingBot-VA, FastWAM, VLA-JEPA, and World Guidance, each produce unsafe actions in every tested category. The result is a benchmark failure signal, not evidence that a training fix is already in place.