Robot manipulation control and safety layers

Manipulation teams working on contact-heavy tasks can add a tactile correction loop around an existing VLA policy before retraining the whole policy. UniTacVLA gives a concrete version: encode current tactile state, predict near-future tactile latents, and let a lightweight Transformer add a bounded residual correction to the planned action at higher frequency than the VLA action chunks.

The cheap validation is a perturbed contact test on a few tasks where vision loses the contact point: USB insertion, wiping, alignment adjustment, or small assembly. UniTacVLA reports 64.0% clean success and 53.5% perturbed success across eight real-robot subtasks, while the reproduced pi0.5-TacVLA baseline reaches 45.25% clean and 16.25% perturbed. Its USB ablation also separates the contribution of tactile input, tactile chain-of-thought supervision, future tactile prediction, and the residual controller, which gives implementers a practical order for testing components.

A robot team with a capable base VLA can test online adaptation at the language-command layer. SARL treats prompts as semantic actions: a controller chooses a short language instruction, the VLA executes it, and reward updates a semantic Q-function over prompts. This is suited to long-horizon failures where the robot has useful primitive skills but the single task prompt leaves it stuck.

A practical pilot would seed three language-steered demonstrations for a task, run online episodes, and compare success under the fixed task prompt against learned prompt sequencing. SARL reports near-zero initial success under the task prompt and about 80% success after 60 to 100 online episodes on Libero-10 and four real WidowX long-horizon tasks. Z-1 adds a related simulation-side check for flow-based VLA policies: after supervised fine-tuning on public RoboCasa demonstrations, GRPO raises average success on 24 RoboCasa tasks from 67.4% to 80.6%.

Household robot evaluators can add a damage score alongside task success before moving a policy into real kitchens or living rooms. OopsieVerse gives a concrete template through DamageSim and OopsieBench: keep per-object and per-robot-link health on a 0 to 100 scale, reduce health when mechanical, thermal, or fluid damage evaluators fire, and report damage separately from completion.

This is most useful for tasks with unsafe shortcuts, such as pushing through fragile objects, contacting hot surfaces, spilling liquid, or overloading robot links. OopsieBench covers 32 task instances across OmniGibson and RoboCasa, and DamageSim is implemented across Nvidia Omniverse and MuJoCo-based stacks. The supplied evidence supports the evaluation workflow more than a policy improvement claim, so the first adoption step is a gate in simulation: pass the task and stay under a damage threshold.