Training changes for reliable long-horizon VLA execution

Teams training long-context policies for multi-stage assembly should regularize the recurrent fast-weight state against the current instruction or sub-task, rather than applying semantic anchoring only to ordinary action features. RoboTTT shows that fast weights can compress 8K timesteps and improve assembly and perturbation recovery, while Semantic Anchoring finds that action–instruction structure erodes during fine-tuning and closely tracks OOD success. Together, they suggest that a larger memory can preserve task-specific shortcuts as readily as useful history.

Add an alignment loss to representations read from the fast weights, with separate channels for semantic state and execution detail. The cheapest diagnostic is to probe action–instruction retrieval after 1K- and 8K-timestep histories, including irrelevant actions and failed attempts; if alignment deteriorates as context grows, compare anchored and unanchored policies on reordered assembly stages and novel task compositions.

VLA teams working on contact, tool-use, and timing tasks should require predicted futures to encode whether a motion path satisfies the task’s physical constraint. FoMoVLA’s combination of a future feature state and sparse point trajectories improves long-horizon manipulation efficiently, but IMBench shows that recognizing constraints or proposing a plan does not imply executable control: GPT-5.5 reached roughly 74% constraint understanding but only 11.3% closed-loop success with vision-only input.

A concrete change is to label predicted point trajectories with constraint outcomes such as maintained contact, clearance, support, tool engagement, or timing, then train the foresight representation to distinguish feasible from visually plausible paths. On IMBench trajectories, compare generic future prediction with this constraint-conditioned objective using both task success and per-category violation rates. This would test whether foresight improves the reasoning-to-action handoff rather than only endpoint prediction.

Training teams that decompose demonstrations into reusable sub-tasks should use those boundary annotations to supervise continuous control across transitions. AC-VLA shows that mixed full-task and sub-task training substantially improves compositional OOD execution, while ChunkFlow shows that inconsistent predictions at shifted chunk boundaries create jitter and accumulated error. The same seam problem can occur where a policy changes from one semantic phase to another, such as grasp to transport or transport to placement.

Retain AC-VLA’s full trajectories, but place overlap windows around aligned sub-task boundaries and apply ChunkFlow-style action, velocity, and curvature consistency losses there. Evaluate OOD recombinations with both success and boundary-local motion metrics; an ablation that moves the same losses away from sub-task transitions would cheaply test whether semantic boundaries are the source of the benefit rather than generic smoothing.