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

VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models

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 VersaCamVLA, which learns a fixed-size scene-token interface from variable posed multi-view RGB via multi-signal target-view prediction and wrist-augmented pose sampling, and injects these tokens into pi0.5 as a supplementary condition for camera-configurable manipulation.

Problem

VLA policies are brittle to changes in camera count or camera pose between training and deployment, since visual token distributions and input view layouts are fixed during action learning.

Approach

Decouple camera-set representation from action learning: a scene encoder trained with RGB, semantic, and edge target-view prediction (LVSM-style) produces fixed-size scene tokens from any posed view subset, and a frozen encoder's compressed tokens are added to a pretrained VLA; WAPS uses wrist-camera motion to diversify training poses.

Key results

• On RoboTwin 2.0 Clean, VersaCamVLA reaches 52.56% average success vs 35.88% for pi0.5 (3 views), and 23.00% on DR vs 21.63% for pi0.5.
• Under unseen camera poses with matched four views, pi0.5 drops 16.0 pp on RoboTwin while VersaCamVLA loses only 0.8 pp (52.63% unseen).
• On LIBERO, VersaCamVLA averages 95.9% vs 95.8% for pi0.5 and 94.5% for OpenVLA-OFT.

Surprising findings

• Adding multi-view tokens directly to pi0.5 degrades performance under unseen poses far more than the scene-token route, and the RoboTwin DR gains are small (23.00 vs 21.63) despite large Clean gains.
• The compressed-token count of 192 outperforms both 48 and 432 tokens, suggesting more tokens do not help.

Limitations

• Assumes calibrated camera intrinsics and extrinsics.
• Focuses on tabletop bimanual setups with manual camera placement; no autonomous viewpoint selection.
• Safety-critical deployment under large camera-configuration shifts is not validated.

Experimental setup

Datasets: RoboTwin 2.0 LIBERO real-robot Cobot Magic demonstrations

Benchmarks: RoboTwin 2.0 LIBERO real-robot Cobot Magic tasks

Metrics: success rate

Baselines: ACT Diffusion Policy DP3 OpenVLA-OFT pi0.5

Builds on: pi0.5 LVSM OpenVLA-OFT SAM 3

Closest papers by meaning

• ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation (NeurIPS 2025 · similarity 0.90)
• BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models (NeurIPS 2025 · similarity 0.89)
• VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.89)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
• Object-Centric Prompt-Driven Vision-Language-Action Model for Robotic Manipulation (CVPR 2025 · similarity 0.88)
• CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.88)