Task-state manipulation stack

Robot teams working on long-horizon table manipulation can now test a planner-executor split with a concrete grounding interface: have a VLM produce the next subtask, target object, and bounding box, then pass both global scene tokens and a high-resolution crop of that box into a separate action policy. HiVLA reports 83.3% average success on RoboTwin 2.0 versus 70.6% for H-RDT, with larger gains on harder tasks and small-object work in clutter. The operational point is narrow and useful: bounding-box-guided local crops give the controller object detail without dropping scene context, while the planner stays out of low-level motor tuning. A cheap validation step is to add planner-produced crops to one cluttered pick-place or tool-use task and compare failure modes against a single-stream policy: missed target identity, poor grasp placement, and loss of multi-step progress are the errors to watch first.

A practical support layer for long-horizon manipulation is explicit step verification and recovery state, not just a stronger action model. Goal2Skill stores episodic history, a compact working-memory summary, and an error register; each subtask carries pre-conditions, post-conditions, horizon, distractor regions, and a primitive skill choice. When execution fails or times out, the system retries, adjusts parameters, or rebuilds the remaining plan. On five RMBench tasks it reports 32.4% average success versus 9.8% for the strongest baseline, and 38.7% versus 9.0% on memory-intensive tasks. For teams already running a VLA executor, this points to a concrete build: add a task-state record and post-condition checks before retraining the policy. The first test is simple: instrument one multi-stage workflow with explicit success checks after each step and measure whether recovery reduces full-task collapse after early mistakes.

Robot RL teams can use a pretrained VLA model as a temporary training signal without turning it into the deployed controller. VLAJS augments PPO with sparse teacher queries on at most 20% of rollout steps, aligns action direction with cosine losses, then decays and removes the guidance as reward improves. The reported result is more than 50% fewer environment interactions on several ManiSkill tasks, while the final policy remains a high-frequency PPO controller. This is a concrete workflow change for labs that already have OpenVLA-class models and expensive simulation or robot time: wire teacher guidance into early training only, and keep deployment on the RL policy. The cheapest check is to run one sparse-reward manipulation task with a fixed query budget and compare time-to-threshold success against PPO and direct distillation baselines.