AnchorVLA: Anchored Diffusion for Efficient End-to-End Mobile Manipulation

Dimensions overall score 7.0

• AnchorVLA reduces inference cost for closed-loop control compared to full iterative denoising diffusion policies.technical 90%
  Evidencepartial

  AnchorVLA combines a lightweight VLA adaptation backbone with an anchored diffusion action head, which denoises locally around anchor trajectories using a truncated diffusion schedule. This retains multimodal action generation while reducing inference cost for closed-loop control.

  Implicationpartial

  Directly stated in abstract as core contribution with clear technical mechanism described

  Verificationpartial

  partial

• AnchorVLA improves success and stability under disturbances and distribution shifts compared to baseline methods.result 85%
  Evidencepartial

  Across diverse mobile manipulation tasks, AnchorVLA improves success and stability under disturbances and distribution shifts while maintaining low-latency inference.

  Implicationpartial

  Directly stated in abstract with implication of experimental validation across diverse tasks

  Verificationpartial

  partial

• Full iterative denoising in diffusion policies is costly at control time for mobile manipulation.limitation 90%
  Evidencepartial

  But in practice, full iterative denoising is costly at control time.

  Implicationpartial

  Directly stated as a problem in the abstract with clear technical reasoning

  Verificationpartial

  partial

• Action chunking in diffusion policies creates partially open-loop behavior that allows small mismatches to accumulate into drift.limitation 85%
  Evidencepartial

  Action chunking helps amortize inference, yet it also creates partially open-loop behavior, allowing small mismatches to accumulate into drift.

  Implicationpartial

  Directly stated as a limitation of existing approaches with clear causal mechanism

  Verificationpartial

  partial

• AnchorVLA introduces a test-time self-correction mechanism via a lightweight residual correction module to mitigate chunking-induced drift.method 90%
  Evidencepartial

  Crucially, to mitigate chunking-induced drift, we introduce a test-time self-correction mechanism via a lightweight residual correction module that makes high-frequency, per-step adjustments during rollout.

  Implicationpartial

  Directly stated as a core technical innovation with specific component description

  Verificationpartial

  partial

• When sampling begins near a plausible solution manifold, extensive denoising is unnecessary to recover multimodal, valid actions.technical 80%
  Evidencepartial

  AnchorVLA, a diffusion-based VLA policy for mobile manipulation built on the core insight that when sampling begins near a plausible solution manifold, extensive denoising is unnecessary to recover multimodal, valid actions.

  Implicationpartial

  Presented as core insight with technical reasoning but requires some inference about implementation

  Verificationpartial

  partial

• AnchorVLA maintains low-latency inference while improving performance on mobile manipulation tasks.result 85%
  Evidencepartial

  Across diverse mobile manipulation tasks, AnchorVLA improves success and stability under disturbances and distribution shifts while maintaining low-latency inference.

  Implicationpartial

  Directly stated in abstract as a key result with clear performance trade-off claim

  Verificationpartial

  partial

• Diffusion policies are appealing for mobile manipulation because they model multimodal action distributions rather than collapsing to one solution.technical 90%
  Evidencepartial

  Diffusion policies are appealing because they model multimodal action distributions rather than collapsing to one solution.

  Implicationpartial

  Directly stated as motivation with clear technical advantage

  Verificationpartial

  partial

Competitive landscape

A diffusion-based policy for mobile manipulation that enables efficient, reactive, and multimodal action generation with self-correction.