Long-horizon robot control is getting more explicit about planning, memory, and guidance

Long-horizon manipulation papers are getting more explicit about the control loop above the actuator policy. HiVLA keeps a vision-language model (VLM) on planning and grounding, then hands execution to a diffusion policy with both global scene tokens and a high-resolution local crop. That setup posts 83.3% average success on RoboTwin 2.0, ahead of H-RDT at 70.6%. Goal2Skill adds structured memory, post-condition checks, and recovery logic for multi-stage tasks. On five RMBench tasks it reports 32.4% average success versus 9.8% for the strongest baseline. The common message is clear: better long-horizon results now come from explicit subtask structure, scene grounding, and step verification, not only from a larger end-to-end policy.

Another strong theme is how to use pretrained vision-language-action priors without locking the robot to the teacher. VLAJS adds sparse guidance from a model such as OpenVLA to PPO, queries the teacher on at most 20% of rollout steps, and aligns action direction while leaving action magnitude to reinforcement learning. The guidance weight decays with reward progress and is removed once learning is stable. In the reported ManiSkill results, this cuts required environment interactions by more than 50% on several tasks, while keeping the final controller as a high-frequency RL policy. This is a practical recipe for borrowing semantic priors early and preserving closed-loop control later.

Aerial language-conditioned robotics appears here as a map of open deployment problems rather than a new model result. The UAV-VLN survey frames drone navigation as a partially observable control problem with natural-language input, then organizes methods across modular systems, spatiotemporal models, and foundation-model agents. Its useful contribution for this period is the constraint list: sim-to-real transfer, outdoor perception drift, ambiguous instructions, and onboard compute remain the main blockers. That makes the paper more valuable as a research agenda than as evidence of benchmark progress.