Embodied AI papers are testing whether robot models can keep control over long execution

Vision-Language-Action (VLA) models are being judged on whether their internal state helps action, prediction, and spatial placement at the same time. Pelican-Unified trains language reasoning, future video generation, and robot action chunks through a shared latent state. It reports 93.5% average success on the 50-task RoboTwin dual-arm suite and an EWM Score of 66.03 on WorldArena.

Evo-Depth attacks a narrower deployment bottleneck: spatial precision without extra depth hardware. Its RGB-derived depth features feed a 0.9B-parameter VLA model. The reported real-world result is 90% average success across three tasks, using 3.2 GB of GPU memory and running at 12.3 Hz.

HandITL focuses on the handoff between an autonomous dexterous policy and a human operator. The key design is relative correction: the operator adjusts the robot hand and arms while the VLA policy continues to run, which avoids a sudden pose jump at takeover.

The measured effect is large in takeover tests. On Bread Clip, direct teleoperation switching caused a mean command change of about 4.38e-2; HandITL reduced it to about 6.8e-5. In Pick Up and Place the Parts, it finished in 42.8 ± 5.0 seconds and had 1 retry across 10 trials. The paper also reports that correction rollouts produced better fine-tuning data than equal-duration standard teleoperation data by 19% on average across three long-horizon dexterous tasks.

LongAct raises the evaluation length for household agents. The benchmark uses 300 episodes in more than 100 ProcTHOR and AI2-THOR houses, with free-form chores averaging about 9 goals and a 16,000-step cap. That setup tests memory, dependency tracking, and recovery across rooms and objects.

The results show a wide gap between goal progress and full completion. Pure Qwen3-VL-32B reaches 6.14% Goal-Condition Success and 0% Success Rate. HoloMind with Qwen3-VL-32B reaches 51.2% Goal-Condition Success and 15.0% Success Rate. The best reported HoloMind variant with GPT-5 reaches 59.0% Goal-Condition Success and 16.0% Success Rate, while human goal completion is reported at 93%.

World-model papers in this window add explicit checks for geometry, object state, and executable dynamics. PDI-Bench scores generated videos on scale-depth alignment, 3D trajectory consistency, and structural rigidity. Ground-truth videos score PDI 0.1206; the best generated model listed, Seedance 2.0, scores 0.2422. Sora and HunyuanVideo show much larger scale residuals, more than 25 times the ground-truth scale residual.

Two planning papers point toward action-facing world models. Slot-MPC uses object slots and a differentiable action-conditioned dynamics model for goal-image manipulation, with a reported 99% latent-size reduction compared with patch-based DINO-WM. Coding Agent Is Good As World Simulator builds PyChrono simulations by writing and repairing code; it reaches 100% Pass@1 plan generation on three reported tasks, but representative successful runs consume millions of tokens.