Robot VLA Work Is Being Judged by Real Control Failure Modes

Scale claims are tied to alignment across robot bodies and sensor choices. Qwen-RobotManip maps different robots into a common state-action template, uses binary masks for missing dimensions, and predicts camera-frame end-effector deltas. Its reported corpus is about 38,100 hours, including synthesized human-to-robot data across 15 bimanual robot configurations, with real-robot validation on AgileX ALOHA, Franka, UR, and ARX.

MuseVLA addresses a different generalization gap. It chooses thermal, acoustic, mmWave, or RGB sensing based on the instruction and scene, then turns the selected signal into a grounded sensor image. On real dexterous-hand tasks, the version with synthesized pretraining reports 80.6% average success on seen tasks and 66.7% on unseen tasks.

Several papers use predictive models during or near execution. DREAM-Chunk samples candidate action chunks from a frozen VLA policy, predicts their latent futures, and chooses actions whose predicted state matches the robot’s observed state. On a precise insertion task under external perturbation, it reports 65% success, compared with 10% for open-loop π0.5.

Mem-World uses a geometry-based wrist-view memory so long rollouts keep object identity and scene layout. Its simulated success estimates correlate with real-world success at r=0.97, and synthetic trajectories help fine-tuned π0.5 reach 72% average success on three long-horizon tasks, compared with a 58% base policy. Qwen-RobotWorld broadens the same idea into language-conditioned video prediction across manipulation, driving, navigation, and human-to-robot data, using 8.6M video-text samples and more than 200M frames.

Action generation is getting more structure. EquiVLA adds SO(2) rotation equivariance to VLA policies built from a frozen vision-language backbone and a flow-matching diffusion transformer action head. The intended behavior is direct: when the scene rotates, the predicted action rotates with it.

The reported gains are concrete. On LIBERO with relative control, EquiVLA reaches 92.6% average success, compared with 78.1% for GR00T N1.5. On five Mobile ALOHA real-robot tasks, it reports 72% average success, compared with 54% for GR00T N1.5. This fits the week’s broader emphasis on action heads, training loops, and predictive models that work on hardware.

Safety work is testing whether embodied models refuse dangerous robot actions before execution. RoboShackles builds 10,000 hazardous robot video clips from real DROID observations, covering hand and human direct harm plus fire, electrical, water, and falling risks.

The strongest result is diagnostic. Six evaluated embodied foundation models each produce unsafe actions in every tested category under the refusal-based criterion. The paper reports no quantified safety-training improvement in the excerpt, so the grounded claim is that current tested systems fail this injury-prevention benchmark.