SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation

TL;DR

SkipVLA is a hybrid manipulation policy that pairs a pretrained VLA with a classical motion planner: the planner handles free-space motion between contact-rich skills, and the VLA is invoked only for grasping/placing. A lightweight target-pose predictor reuses the frozen VLA’s vision-language backbone to emit each planner subgoal, learned without additional demonstrations by distilling what the large VLA already knows. Across three VLAs on 13 LIBERO tasks and three pick-and-place tasks on a physical 6-DoF YAM arm, SkipVLA reports up to 2.5× faster task completion and lower energy consumption at matched task success.

Key claims

• Free-space motion segments of long-horizon manipulation tasks can be handed to a classical motion planner while a pretrained VLA is invoked only for contact-rich skills such as grasping and placing, without loss of task success [Abstract].
• Target poses for each planned motion can be predicted by reusing the frozen VLA’s own vision-language backbone, and this predictor is trained without any additional demonstrations beyond what the underlying VLA was trained on [Abstract].
• On 13 LIBERO tasks in simulation across three different VLAs, SkipVLA delivers up to 2.5× faster task completion at matched task success [Abstract].
• On three real pick-and-place tasks on a 6-DoF YAM arm, SkipVLA reports the same task success as the underlying VLA with significantly lower energy consumption [Abstract].

Method

SkipVLA is a segmentation-based hybrid policy: given a language instruction and camera images, it decides which segments of the task are free-space transit versus contact-rich manipulation, calls a classical collision-free motion planner for the former, and hands control to a pretrained VLA for the latter. The linkage between the two is a target-pose predictor that reads the frozen VLA’s vision-language backbone features and emits an SE(3) target pose per planned motion; this predictor is trained to reproduce the target poses implicit in the VLA’s own action rollouts, so no new demonstration data is introduced. Three VLA backbones are exercised in the LIBERO evaluation; the physical evaluation runs on a 6-DoF YAM arm.

Results

• Simulation (LIBERO, 13 tasks, three VLA backbones): up to 2.5× faster task completion at matched task success rate [Abstract].
• Real hardware (6-DoF YAM arm, three pick-and-place tasks): matched task success rate vs the underlying VLA, with significantly lower energy consumption [Abstract].

Per-task and per-backbone breakdowns are not extractable from the abstract at filing time.

Why it’s interesting

SkipVLA occupies a distinct row on the VLA Models recipe-lever board: instead of accelerating the VLA itself (quantized serving à la APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin, sparse caching à la C³ache: Accelerating World Action Models with Cross Inference Chunk Cache, LLM-free architectures à la TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM) or restructuring async chunking (πR² πR²: Reactive Real-time Flow Policies, A2A, REMAC Real-Time Robot Execution with Masked Action Chunking), it skips VLA calls entirely on free-space segments — the classical-planner counterpart to Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents Harness VLA’s “VLA as narrow retryable contact-rich primitive” thesis, but with a learned target-pose predictor as the planner-to-VLA glue rather than a memory-augmented LLM planner. Complements the LLM Inference Efficiency page’s async / caching / quantization axes with a scheduling axis: don’t just make each VLA call cheaper, make fewer of them. The zero-additional-demonstrations training story (bootstrap the pose predictor off the underlying VLA) also puts it adjacent to policy-distillation moves like Proxy Policy Steering: Adapting Frozen Robotic Foundation Models via Velocity-Space Proxy Residuals Proxy Policy Steering, where a lightweight learned module wraps a frozen base without touching its weights or demanding new labels.

See also

• VLA Models — SkipVLA sits on the recipe-lever board as a hybrid-with-classical-planner row; complementary to VLA-only levers (RL, clean teleop, unified-VLM pointing, WFM+action-expert).
• LLM Inference Efficiency — VLA-latency reduction via call-count reduction rather than per-call speedup.
• Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents — Harness VLA also restricts a frozen VLA to contact-rich primitives, but wraps it in an LLM planner + analytic primitives rather than a classical motion planner.
• ModPack: An Extensible Teleoperation Interface for Bimanual Mobile Manipulation — ModPack contains a related bimanual data / policy stack that also isolates contact-critical phases.
• APXinf-robo — Edge Inference Engine for VLA/WAM Models on Jetson Thor / Orin — an orthogonal per-call VLA-latency axis (Rust + CUTLASS + FlashAttention Jetson serving) that would compose with call-count reduction.
• C³ache: Accelerating World Action Models with Cross Inference Chunk Cache — cross-chunk residual caching, another VLA-latency lever that composes with SkipVLA’s segmentation approach.
• TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM — LLM-free VLA architecture as an alternative latency lever.