Scaling Population-Based Reinforcement Learning with GPU Accelerated Simulation

In recent years, deep reinforcement learning (RL) has shown its effectiveness in solving complex continuous control tasks like locomotion and dexterous manipulation. However, this comes at the cost of an enormous amount of experience required for training, exacerbated by the sensitivity of learning efficiency and the policy performance to hyperparameter selection, which often requires numerous trials of time-consuming experiments. This work introduces a Population-Based Reinforcement Learning (PBRL) approach that exploits a GPU-accelerated physics simulator to enhance the exploration capabilities of RL by concurrently training multiple policies in parallel. The PBRL framework is applied to three state-of-the-art RL algorithms -- PPO, SAC, and DDPG -- dynamically adjusting hyperparameters based on the performance of learning agents. The experiments are performed on four challenging tasks in Isaac Gym -- Anymal Terrain, Shadow Hand, Humanoid, Franka Nut Pick -- by analyzing the effect of population size and mutation mechanisms for hyperparameters. The results show that PBRL agents achieve superior performance, in terms of cumulative reward, compared to non-evolutionary baseline agents. The trained agents are finally deployed in the real world for a Franka Nut Pick task, demonstrating successful sim-to-real transfer. Code and videos of the learned policies are available on our project website.

Perfecting Periodic Trajectory Tracking: Model Predictive Control with a Periodic Observer ($Π$-MPC)

In Model Predictive Control (MPC), discrepancies between the actual system and the predictive model can lead to substantial tracking errors and significantly degrade performance and reliability. While such discrepancies can be alleviated with more complex models, this often complicates controller design and implementation. By leveraging the fact that many trajectories of interest are periodic, we show that perfect tracking is possible when incorporating a simple observer that estimates and compensates for periodic disturbances. We present the design of the observer and the accompanying tracking MPC scheme, proving that their combination achieves zero tracking error asymptotically, regardless of the complexity of the unmodelled dynamics. We validate the effectiveness of our method, demonstrating asymptotically perfect tracking on a high-dimensional soft robot with nearly 10,000 states and a fivefold reduction in tracking errors compared to a baseline MPC on small-scale autonomous race car experiments.

InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds

While novel view synthesis (NVS) has made substantial progress in 3D computer vision, it typically requires an initial estimation of camera intrinsics and extrinsics from dense viewpoints. This pre-processing is usually conducted via a Structure-from-Motion (SfM) pipeline, a procedure that can be slow and unreliable, particularly in sparse-view scenarios with insufficient matched features for accurate reconstruction. In this work, we integrate the strengths of point-based representations (e.g., 3D Gaussian Splatting, 3D-GS) with end-to-end dense stereo models (DUSt3R) to tackle the complex yet unresolved issues in NVS under unconstrained settings, which encompasses pose-free and sparse view challenges. Our framework, InstantSplat, unifies dense stereo priors with 3D-GS to build 3D Gaussians of large-scale scenes from sparseview & pose-free images in less than 1 minute. Specifically, InstantSplat comprises a Coarse Geometric Initialization (CGI) module that swiftly establishes a preliminary scene structure and camera parameters across all training views, utilizing globally-aligned 3D point maps derived from a pre-trained dense stereo pipeline. This is followed by the Fast 3D-Gaussian Optimization (F-3DGO) module, which jointly optimizes the 3D Gaussian attributes and the initialized poses with pose regularization. Experiments conducted on the large-scale outdoor Tanks & Temples datasets demonstrate that InstantSplat significantly improves SSIM (by 32%) while concurrently reducing Absolute Trajectory Error (ATE) by 80%. These establish InstantSplat as a viable solution for scenarios involving posefree and sparse-view conditions. Project page: instantsplat.github.io.

Producing and Leveraging Online Map Uncertainty in Trajectory Prediction

High-definition (HD) maps have played an integral role in the development of modern autonomous vehicle (AV) stacks, albeit with high associated labeling and maintenance costs. As a result, many recent works have proposed methods for estimating HD maps online from sensor data, enabling AVs to operate outside of previously-mapped regions. However, current online map estimation approaches are developed in isolation of their downstream tasks, complicating their integration in AV stacks. In particular, they do not produce uncertainty or confidence estimates. In this work, we extend multiple state-of-the-art online map estimation methods to additionally estimate uncertainty and show how this enables more tightly integrating online mapping with trajectory forecasting. In doing so, we find that incorporating uncertainty yields up to 50% faster training convergence and up to 15% better prediction performance on the real-world nuScenes driving dataset.

Task-Driven Manipulation with Reconfigurable Parallel Robots

ReachBot, a proposed robotic platform, employs extendable booms as limbs for mobility in challenging environments, such as martian caves. When attached to the environment, ReachBot acts as a parallel robot, with reconfiguration driven by the ability to detach and re-place the booms. This ability enables manipulation-focused scientific objectives: for instance, through operating tools, or handling and transporting samples. To achieve these capabilities, we develop a two-part solution, optimizing for robustness against task uncertainty and stochastic failure modes. First, we present a mixed-integer stance planner to determine the positioning of ReachBot's booms to maximize the task wrench space about the nominal point(s). Second, we present a convex tension planner to determine boom tensions for the desired task wrenches, accounting for the probabilistic nature of microspine grasping. We demonstrate improvements in key robustness metrics from the field of dexterous manipulation, and show a large increase in the volume of the manipulation workspace. Finally, we employ Monte-Carlo simulation to validate the robustness of these methods, demonstrating good performance across a range of randomized tasks and environments, and generalization to cable-driven morphologies. We make our code available at our project webpage, https://stanfordasl.github.io/reachbot_manipulation/

Q-SLAM: Quadric Representations for Monocular SLAM

Monocular SLAM has long grappled with the challenge of accurately modeling 3D geometries. Recent advances in Neural Radiance Fields (NeRF)-based monocular SLAM have shown promise, yet these methods typically focus on novel view synthesis rather than precise 3D geometry modeling. This focus results in a significant disconnect between NeRF applications, i.e., novel-view synthesis and the requirements of SLAM. We identify that the gap results from the volumetric representations used in NeRF, which are often dense and noisy. In this study, we propose a novel approach that reimagines volumetric representations through the lens of quadric forms. We posit that most scene components can be effectively represented as quadric planes. Leveraging this assumption, we reshape the volumetric representations with million of cubes by several quadric planes, which leads to more accurate and efficient modeling of 3D scenes in SLAM contexts. Our method involves two key steps: First, we use the quadric assumption to enhance coarse depth estimations obtained from tracking modules, e.g., Droid-SLAM. This step alone significantly improves depth estimation accuracy. Second, in the subsequent mapping phase, we diverge from previous NeRF-based SLAM methods that distribute sampling points across the entire volume space. Instead, we concentrate sampling points around quadric planes and aggregate them using a novel quadric-decomposed Transformer. Additionally, we introduce an end-to-end joint optimization strategy that synchronizes pose estimation with 3D reconstruction.

Mapping High-level Semantic Regions in Indoor Environments without Object Recognition

Robots require a semantic understanding of their surroundings to operate in an efficient and explainable way in human environments. In the literature, there has been an extensive focus on object labeling and exhaustive scene graph generation; less effort has been focused on the task of purely identifying and mapping large semantic regions. The present work proposes a method for semantic region mapping via embodied navigation in indoor environments, generating a high-level representation of the knowledge of the agent. To enable region identification, the method uses a vision-to-language model to provide scene information for mapping. By projecting egocentric scene understanding into the global frame, the proposed method generates a semantic map as a distribution over possible region labels at each location. This mapping procedure is paired with a trained navigation policy to enable autonomous map generation. The proposed method significantly outperforms a variety of baselines, including an object-based system and a pretrained scene classifier, in experiments in a photorealistic simulator.

Parallelized Spatiotemporal Binding

While modern best practices advocate for scalable architectures that support long-range interactions, object-centric models are yet to fully embrace these architectures. In particular, existing object-centric models for handling sequential inputs, due to their reliance on RNN-based implementation, show poor stability and capacity and are slow to train on long sequences. We introduce Parallelizable Spatiotemporal Binder or PSB, the first temporally-parallelizable slot learning architecture for sequential inputs. Unlike conventional RNN-based approaches, PSB produces object-centric representations, known as slots, for all time-steps in parallel. This is achieved by refining the initial slots across all time-steps through a fixed number of layers equipped with causal attention. By capitalizing on the parallelism induced by our architecture, the proposed model exhibits a significant boost in efficiency. In experiments, we test PSB extensively as an encoder within an auto-encoding framework paired with a wide variety of decoder options. Compared to the state-of-the-art, our architecture demonstrates stable training on longer sequences, achieves parallelization that results in a 60% increase in training speed, and yields performance that is on par with or better on unsupervised 2D and 3D object-centric scene decomposition and understanding.

Contingency Planning Using Bi-level Markov Decision Processes for Space Missions

This work focuses on autonomous contingency planning for scientific missions by enabling rapid policy computation from any off-nominal point in the state space in the event of a delay or deviation from the nominal mission plan. Successful contingency planning involves managing risks and rewards, often probabilistically associated with actions, in stochastic scenarios. Markov Decision Processes (MDPs) are used to mathematically model decision-making in such scenarios. However, in the specific case of planetary rover traverse planning, the vast action space and long planning time horizon pose computational challenges. A bi-level MDP framework is proposed to improve computational tractability, while also aligning with existing mission planning practices and enhancing explainability and trustworthiness of AI-driven solutions. We discuss the conversion of a mission planning MDP into a bi-level MDP, and test the framework on RoverGridWorld, a modified GridWorld environment for rover mission planning. We demonstrate the computational tractability and near-optimal policies achievable with the bi-level MDP approach, highlighting the trade-offs between compute time and policy optimality as the problem's complexity grows. This work facilitates more efficient and flexible contingency planning in the context of scientific missions.

Driving Everywhere with Large Language Model Policy Adaptation

Adapting driving behavior to new environments, customs, and laws is a long-standing problem in autonomous driving, precluding the widespread deployment of autonomous vehicles (AVs). In this paper, we present LLaDA, a simple yet powerful tool that enables human drivers and autonomous vehicles alike to drive everywhere by adapting their tasks and motion plans to traffic rules in new locations. LLaDA achieves this by leveraging the impressive zero-shot generalizability of large language models (LLMs) in interpreting the traffic rules in the local driver handbook. Through an extensive user study, we show that LLaDA's instructions are useful in disambiguating in-the-wild unexpected situations. We also demonstrate LLaDA's ability to adapt AV motion planning policies in real-world datasets; LLaDA outperforms baseline planning approaches on all our metrics. Please check our website for more details: https://boyiliee.github.io/llada.