RSS 2023 Papers — Page 2
Robotics: Science and Systems · 112 papers
Sequence-Based Plan Feasibility Prediction for Efficient Task and Motion Planning
Zhutian Yang (Massachusetts Institute of Technology), Dieter Fox (NVIDIA Research)
OptimizationRobotic IntelligenceTransformerSupervised Fine-TuningImageTextSequential
🎯 What it does: Propose a learning-driven task and motion planning (TAMP) framework that accelerates robot rearrangement tasks in kitchen environments with movable and articulated objects by predicting the feasibility of task plans.
Simultaneous Trajectory Optimization and Contact Selection for Multi-Modal Manipulation Planning
Mengchao Zhang (University of Illinois Urbana-Champaign), Kris Hauser (University of Illinois Urbana-Champaign)
OptimizationRobotic Intelligence
🎯 What it does: Propose an algorithm called STOCS that simultaneously performs trajectory optimization and contact selection for multimodal manipulation planning, and embeds it into a sampling-based tree search to enable automatic planning of multiple contact modes such as grasping, pushing/swinging, and rotating.
Solving Stabilize-Avoid via Epigraph Form Optimal Control using Deep Reinforcement Learning
Oswin So (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
OptimizationRobotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: This paper proposes solving the robot's stabilize-avoid control tasks by converting the infinite-horizon constrained optimal control problem into an epigraph form and using deep reinforcement learning (EFPPO).
StructDiffusion: Language-Guided Creation of Physically-Valid Structures using Unseen Objects
Weiyu Liu (Georgia Tech), Chris Paxton (Meta AI)
GenerationPose EstimationRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelGenerative Adversarial NetworkTextPoint Cloud
🎯 What it does: This paper proposes a framework called StructDiffusion, which can generate physically feasible and semantically constrained multi-object structures for unknown objects under conditions where only partial point cloud perspectives and natural language instructions are available. It samples diverse target poses and then enables the robot to perform pick-and-place operations to rearrange objects.
Structured World Models from Human Videos
Russell Mendonca (Carnegie Mellon University), Deepak Pathak
Robotic IntelligenceConvolutional Neural NetworkReinforcement LearningVision-Language-Action ModelWorld ModelImageVideo
🎯 What it does: A framework that efficiently learns multiple manipulation tasks within 30 minutes was achieved by pre-training a structured world model on large-scale human videos and then fine-tuning it on a small amount of robot interaction data.
Tactile-Filter: Interactive Tactile Perception for Part Mating
Kei Ota (Mitsubishi Electric), Joshua Tenenbaum
RecognitionRobotic IntelligenceContrastive LearningImage
🎯 What it does: This paper proposes an interactive perception method called Tactile-Filter based on visual tactile sensors, which identifies and locates mating parts with a limited number of touches without prior 3D models.
Task-Aware Risk Estimation of Perception Failures for Autonomous Vehicles
Pasquale Antonante (Massachusetts Institute of Technology), Marco Pavone (NVIDIA)
Autonomous DrivingMultimodality
🎯 What it does: Propose a task-aware risk estimation framework to assess the risk of perception failure on autonomous vehicle motion planning and provide a decision-making algorithm for triggering safety actions.
Teach a Robot to FISH: Versatile Imitation from One Minute of Demonstrations
Siddhant Haldar (New York University), Lerrel Pinto (New York University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningVideo
🎯 What it does: This paper proposes a fast imitation learning method called FISH, which uses less than one minute of demonstration data. It first trains a weak baseline policy and then fine-tunes a residual policy through online reinforcement learning (RL), enabling robots to complete tasks under unseen object configurations.
TerrainNet: Visual Modeling of Complex Terrain for High-speed, Off-road Navigation
Xiangyun Meng (University of Washington), Amirreza Shaban (University of Washington)
Depth EstimationAutonomous DrivingOptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningSimultaneous Localization and MappingImageVideoPoint Cloud
🎯 What it does: The paper proposes TerrainNet, a real-time vision-based terrain perception system using cameras, capable of generating multi-layer bird's-eye-view maps containing semantic and geometric information for assessing terrain traversability and driving planning in high-speed off-road driving.
Time Optimal Ergodic Search
Dayi E Dong (Yale University), Ian Abraham (Yale University)
OptimizationRobotic IntelligenceTime Series
🎯 What it does: Propose a time-optimal Ergodic search method that generates coverage trajectories in the shortest time under a given information distribution, balancing search quality and time cost.
To the Noise and Back: Diffusion for Shared Autonomy
Takuma Yoneda (Toyota Technological Institute at Chicago), Matthew R Walter
Robotic IntelligenceDiffusion modelSequential
🎯 What it does: This paper proposes a shared autonomy method based on diffusion models, which maps user actions to a trained target behavior distribution by utilizing partial forward and backward diffusion processes, thereby achieving human-machine collaborative control without requiring environmental dynamics, target spaces, or reward information.
Uncertain Pose Estimation during Contact Tasks using Differentiable Contact Features
Dongjun Lee (Seoul National University), Minji Lee (Seoul National University)
Pose EstimationRobotic IntelligenceImagePoint Cloud
🎯 What it does: Propose a differentiable framework that uses gradient optimization methods to accurately estimate uncertain poses in contact tasks.