Research map / Vision-language and multimodal / Robot action and navigation

PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PhaseLoRA, a LoRA variant whose left factor is modulated per action-chunk step by router-predicted fine-control and event/boundary descriptors, evaluated on LIBERO with pi0.5.

Problem

Standard PEFT/LoRA adapters for continuous-action VLA policies apply a temporally static update across a rollout, ignoring that manipulation passes through distinct control regimes (approach, grasp, transport, placement).

Approach

Condition the LoRA left factor on two weakly supervised, action-derived descriptors (fine-control tendency P and event/boundary intensity E) via B(P,E)=B0+P·BP+E·BE+PE·BPE, so the effective low-rank update direction changes within a rollout while the backbone stays frozen.

Key results

• On LIBERO, PhaseLoRA reaches 68.9% average success vs 56.7% for matched-parameter High-rank LoRA (+12.2 points) and 38.3% for standard LoRA.
• On the real-world Piper benchmark, PhaseLoRA averages 69.2% vs 50.8% for High-rank LoRA (+18.4 points).
• On LIBERO-Spatial ablations, random descriptors drop to 66.3% and scalar-gated update to 73.7%, versus 85.3% for the full model.

Surprising findings

• Removing the interaction term B_PE yields only a small drop (85.3 to 82.3), suggesting most gains come from additive descriptor conditioning.
• Scalar-gated LoRA with the same router performs only slightly above High-rank LoRA on LIBERO-Spatial (73.7 vs 71.0).

Limitations

• Descriptors are action-derived weak proxies, which may not reflect fine-control or boundary events in task families where action statistics are uninformative.
• Detailed ablations only on LIBERO-Spatial; real-world evaluation limited to four tasks on one robot platform.
• Real-world success judged manually; results reported with only 3 seeds.

Experimental setup

Datasets: LIBERO-Spatial LIBERO-Object LIBERO-Goal LIBERO-10 Real-world Piper tabletop demonstrations (4 tasks, 100 each)

Benchmarks: LIBERO-Spatial LIBERO-Object LIBERO-Goal LIBERO-10 Real-world Piper tabletop (4 tasks)

Metrics: success rate trainable parameter percentage Spearman correlation update-direction cosine distance

Baselines: LoRA High-rank LoRA DoRA LoRA-MoE LoRA-SP Scalar-gated LoRA

Builds on: LoRA DoRA pi0.5 OpenVLA-OFT LoRA-SP

Closest papers by meaning

• PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.89)
• VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.88)
• FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.88)