Most vision–language–action (VLA) models predict actions but don’t by themselves make reliable long-horizon agents. The core insight behind this work is to treat each high-level skill choice as an execution proposal that must pass runtime preflight checks and post-execution verification, transforming black-box action generation into a traceable, trainable closed-loop agent layer.
Key Findings
- Modular skill interface: skills have typed inputs/outputs and explicit prerequisites, so the runtime can block invalid invocations before physical actions occur — this reduces silent failures and makes trajectories interpretable for supervision.
- Strong cross-benchmark execution: task-adapted low-level VLA policies in the paper achieve 86.20% average success over 50 RoboTwin 2.0 tasks and 97.40% across four LIBERO suites, demonstrating robust short-horizon execution when integrated into the agent loop.
- Trajectory-driven training: the system records multimodal context, runtime errors, and verifier outputs so planners, low-level policies, and verifiers can be trained or adapted independently using subtask-level supervision or optional online feedback.
- Clear limits revealed: the same execution approach yields only 12.5% average success on memory-dependent RMBench tasks, highlighting that skill orchestration improves deployment reliability but does not alone solve long-term memory or complex reasoning deficits.
Who It's For and Trade-offs
Great fit if you need to convert VLA models into inspectable, deployable embodied systems where runtime safety checks, outcome verification, and per-component training matter. It’s especially useful when you want to swap or fine-tune low-level execution policies without changing orchestration logic. Look elsewhere if your primary bottleneck is long-term memory, complex multi-step reasoning beyond short-horizon execution chunks, or when you require end-to-end learned policies with no modular runtime checks.
Where It Fits
EmbodiedSkills sits between high-level planners and robot controllers: it does not replace low-level VLA action models but wraps them with pre-execution validation, bounded execution windows, post-execution verifiers, and structured logging so that embodied agents become safer, debuggable, and easier to train component-wise.