Research map / Vision-language and multimodal / Robot action and navigation
OASIS: Observation-Action Space Alignment via SE(3) Trajectory Prediction for Robotic Manipulation
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 OASIS, a visuomotor policy that predicts a camera-frame SE(3) end-effector trajectory from 3D-aware (VLM + metric depth) features and conditions an action decoder on its pose-supervised hidden states, reporting strong LIBERO, CALVIN, and real-robot results.
Problem
Existing VLA models and world action models decode 6-DoF actions from intermediate representations (depth/spatial features, future images, latent visual features) that do not share the rigid-body geometry of the action space, forcing the decoder to implicitly recover pose.
Approach
Insert an SE(3) camera-frame end-effector trajectory predictor between a 3D-aware encoder and the action decoder, and condition the decoder on the predictor's pose-supervised hidden states so the intermediate is geometrically aligned with the action space, trained end-to-end with only standard demonstrations.
Key results
- OASIS achieves 97.6% average success on LIBERO across four suites, versus 95.5% for Unified-VLA (best baseline in table).
- On CALVIN ABC->D, OASIS reaches 4.57 average sequence length and 83.3% five-task success.
- Real-world average success of 89.2% across Goal, Spatial, and Long suites, vs 81.4%-ish for pi0.5 (7.6 points lower per text).
Surprising findings
- A hardcoded geometric pipeline reading the same predicted SE(3) trajectory collapses to 12.4% on LIBERO-Spatial and 0.0% on LIBERO-Long, showing the learned decoder is necessary.
- Predicting in the world frame underperforms camera-frame prediction by 2.0 points on LIBERO-Long.
Limitations
- Focuses on tabletop manipulation with a single arm; does not address mobile or dexterous action spaces.
- Real-world Long task gap over pi0.5 is within overlapping Wilson intervals, so not statistically clear.
- OOD camera test only displaces the third-person camera by about 15 cm, not full extrinsic invariance.
Experimental setup
Datasets: LIBERO CALVIN LLaVA-1.5-Instruct
Benchmarks: LIBERO CALVIN ABC->D real-world Franka Research 3 / Kinova Gen3 suites (Goal, Spatial, Long)
Metrics: success rate average sequence length tasks completed in a row Wilson 95% confidence interval
Baselines: SpatialVLA WorldVLA ThinkAct pi0 QDepth-VLA UniVLA Unified-VLA SuSIE 3D Diffuser Actor ReconVLA Seer-Large VPP
Builds on: Depth Anything 3 Prismatic VLM Seer LoRA
Closest papers by meaning
- Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.88)
- PIVOT-R: Primitive-Driven Waypoint-Aware World Model for Robotic Manipulation (NeurIPS 2024 · similarity 0.88)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.88)
- PDFactor: Learning Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic Manipulation (CVPR 2025 · similarity 0.88)
- Localizing, Structuring, and Rendering: Bridging 3D and 2D Vision-Language-Action Models for Robotic Manipulation (CVPR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.