Robot papers are concentrating on temporal memory, recovery, and usable control loops

Several papers treat time as a first-class input to robot control. MemoryVLA++ stores past perceptual and task tokens, then combines them with latent future prediction before action generation. Its reported gains are largest on tasks that need remembering a prior interaction or anticipating motion, including +26 and +28 percentage points on real-robot memory- and imagination-dependent task groups.

iMaC uses robot kinematics and contact heatmaps to make action-conditioned video rollouts more spatially precise. Its world-model estimates correlate with real policy success on six of eight long-horizon real-robot tasks. Echo-Memory adds a useful warning: replay metrics alone can miss whether a world model preserves object identity after the camera leaves and returns.

Recovery work keeps the base VLA fixed and adds targeted correction at runtime or through residual training. ReCoVLA uses a vision-language model to identify failure type and recovery stage, compiles stage-gated rewards, and trains residual policies in simulation. It raises average Fetch-task simulation success from 36.7% to 66.7%, with zero-shot physical success at 61.7%.

B2FF pre-generates future-image milestones and selects a recovery goal when execution drifts. On failure-injected LIBERO, it improves UD-VLA average success from 56.3% to 74.0%. ProbeAct is lighter weight: it reads hidden-state probes to track objects and uses control-barrier zones after repeated failures, improving LIBERO-plus success from 69.6% to 74.1%.

The execution loop is getting more explicit user and safety channels. Flow Control modifies the initial condition of a flow-matching action sampler with a simple keyboard direction. In the Two-Block task, steering the initial condition made left-block acquisition nearly perfect at longer steering horizons while pick-and-place success stayed near 100% in the reported setup.

A separate safety-filter paper reads attention heads inside a frozen VLA to identify the current target object during control. On dynamic SafeLIBERO Level III, it cuts average collision rate from 70.75% for an init-only filter to 26.88%, and raises safe-success from 25.5% to 55.75%. The common point is practical: small runtime signals can constrain a large policy without retraining it.

YUBI attacks the data bottleneck with a lighter finger-aligned gripper and fixed-rig VR tracking. The reported dataset is unusually large for an open UMI-style setup: 8,434 hours, 1.20 million episodes, 6.80 million video-language-action triplets, and 119 tasks collected by 179 operators. A single π0.5-based policy trained on YUBI wrist data transfers across UR, Franka, and Toyota ELEY arms with the same end effector.

MotionWAM extends world-action modeling to Unitree G1 humanoid loco-manipulation. It uses one egocentric camera and unified whole-body motion latents, reaching 76.1% average success across nine real tasks versus 43.9% for the strongest listed baseline. TORL-VLA adds tactile online reinforcement learning for contact-rich tasks, reaching 30/30 cup, 29/30 latch, and 30/30 egg successes in real-robot subtask trials.