Inference-Time Policy Steering through Human Interactions

Generative policies trained with human demonstrations can autonomously accomplish multimodal, long-horizon tasks. However, during inference, humans are often removed from the policy execution loop, limiting the ability to guide a pre-trained policy towards a specific sub-goal or trajectory shape among multiple predictions. Naive human intervention may inadvertently exacerbate distribution shift, leading to constraint violations or execution failures. To better align policy output with human intent without inducing out-of-distribution errors, we propose an Inference-Time Policy Steering (ITPS) framework that leverages human interactions to bias the generative sampling process, rather than fine-tuning the policy on interaction data. We evaluate ITPS across three simulated and real-world benchmarks, testing three forms of human interaction and associated alignment distance metrics. Among six sampling strategies, our proposed stochastic sampling with diffusion policy achieves the best trade-off between alignment and distribution shift. Videos are available at https://yanweiw.github.io/itps/.

First Place Solution to the ECCV 2024 ROAD++ Challenge @ ROAD++ Atomic Activity Recognition 2024

This report presents our team's technical solution for participating in Track 3 of the 2024 ECCV ROAD++ Challenge. The task of Track 3 is atomic activity recognition, which aims to identify 64 types of atomic activities in road scenes based on video content. Our approach primarily addresses the challenges of small objects, discriminating between single object and a group of objects, as well as model overfitting in this task. Firstly, we construct a multi-branch activity recognition framework that not only separates different object categories but also the tasks of single object and object group recognition, thereby enhancing recognition accuracy. Subsequently, we develop various model ensembling strategies, including integrations of multiple frame sampling sequences, different frame sampling sequence lengths, multiple training epochs, and different backbone networks. Furthermore, we propose an atomic activity recognition data augmentation method, which greatly expands the sample space by flipping video frames and road topology, effectively mitigating model overfitting. Our methods rank first in the test set of Track 3 for the ROAD++ Challenge 2024, and achieve 69% mAP.

Versatile Demonstration Interface: Toward More Flexible Robot Demonstration Collection

Previous methods for Learning from Demonstration leverage several approaches for a human to teach motions to a robot, including teleoperation, kinesthetic teaching, and natural demonstrations. However, little previous work has explored more general interfaces that allow for multiple demonstration types. Given the varied preferences of human demonstrators and task characteristics, a flexible tool that enables multiple demonstration types could be crucial for broader robot skill training. In this work, we propose Versatile Demonstration Interface (VDI), an attachment for collaborative robots that simplifies the collection of three common types of demonstrations. Designed for flexible deployment in industrial settings, our tool requires no additional instrumentation of the environment. Our prototype interface captures human demonstrations through a combination of vision, force sensing, and state tracking (e.g., through the robot proprioception or AprilTag tracking). Through a user study where we deployed our prototype VDI at a local manufacturing innovation center with manufacturing experts, we demonstrated the efficacy of our prototype in representative industrial tasks. Interactions from our study exposed a range of industrial use cases for VDI, clear relationships between demonstration preferences and task criteria, and insights for future tool design.

Grounding Language Plans in Demonstrations Through Counterfactual Perturbations

Grounding the common-sense reasoning of Large Language Models in physical domains remains a pivotal yet unsolved problem for embodied AI. Whereas prior works have focused on leveraging LLMs directly for planning in symbolic spaces, this work uses LLMs to guide the search of task structures and constraints implicit in multi-step demonstrations. Specifically, we borrow from manipulation planning literature the concept of mode families, which group robot configurations by specific motion constraints, to serve as an abstraction layer between the high-level language representations of an LLM and the low-level physical trajectories of a robot. By replaying a few human demonstrations with synthetic perturbations, we generate coverage over the demonstrations' state space with additional successful executions as well as counterfactuals that fail the task. Our explanation-based learning framework trains an end-to-end differentiable neural network to predict successful trajectories from failures and as a by-product learns classifiers that ground low-level states and images in mode families without dense labeling. The learned grounding classifiers can further be used to translate language plans into reactive policies in the physical domain in an interpretable manner. We show our approach improves the interpretability and reactivity of imitation learning through 2D navigation and simulated and real robot manipulation tasks. Website: https://sites.google.com/view/grounding-plans

Human-Machine Cooperative Multimodal Learning Method for Cross-subject Olfactory Preference Recognition

Odor sensory evaluation has a broad application in food, clothing, cosmetics, and other fields. Traditional artificial sensory evaluation has poor repeatability, and the machine olfaction represented by the electronic nose (E-nose) is difficult to reflect human feelings. Olfactory electroencephalogram (EEG) contains odor and individual features associated with human olfactory preference, which has unique advantages in odor sensory evaluation. However, the difficulty of cross-subject olfactory EEG recognition greatly limits its application. It is worth noting that E-nose and olfactory EEG are more advantageous in representing odor information and individual emotions, respectively. In this paper, an E-nose and olfactory EEG multimodal learning method is proposed for cross-subject olfactory preference recognition. Firstly, the olfactory EEG and E-nose multimodal data acquisition and preprocessing paradigms are established. Secondly, a complementary multimodal data mining strategy is proposed to effectively mine the common features of multimodal data representing odor information and the individual features in olfactory EEG representing individual emotional information. Finally, the cross-subject olfactory preference recognition is achieved in 24 subjects by fusing the extracted common and individual features, and the recognition effect is superior to the state-of-the-art recognition methods. Furthermore, the advantages of the proposed method in cross-subject olfactory preference recognition indicate its potential for practical odor evaluation applications.

Dr. LLaMA: Improving Small Language Models on PubMedQA via Generative Data Augmentation

Large Language Models (LLMs) have made remarkable strides in natural language processing, but their expanding size poses challenges in terms of computational expense and inefficiency. Conversely, Small Language Models (SLMs) are known for their efficiency but often encounter difficulties in tasks with limited capacity and training data, particularly in domain-specific scenarios. In this paper, we introduce Dr. LLaMA, a method that improves SLMs in the medical domain through generative data augmentation utilizing LLMs. The objective is to develop more efficient and capable models tailored for specialized applications. Our preliminary results on the PubMedQA dataset demonstrate that LLMs effectively refine and diversify existing question-answer pairs, leading to improved performance of a significantly smaller model after fine-tuning. The best SLM surpasses few-shot GPT-4 with under 1.6 billion parameters on the PubMedQA. Our code and generated data are publicly available to facilitate further explorations.

Visual Pre-training for Navigation: What Can We Learn from Noise?

A powerful paradigm for sensorimotor control is to predict actions from observations directly. Training such an end-to-end system allows representations that are useful for the downstream tasks to emerge automatically. In visual navigation, an agent can learn to navigate without any manual designs by correlating how its views change with the actions being taken. However, the lack of inductive bias makes this system data-inefficient and impractical in scenarios like search and rescue, where interacting with the environment to collect data is costly. We hypothesize a sufficient representation of the current view and the goal view for a navigation policy can be learned by predicting the location and size of a crop of the current view that corresponds to the goal. We further show that training such random crop prediction in a self-supervised fashion purely on random noise images transfers well to natural home images. The learned representation can then be bootstrapped to learn a navigation policy efficiently with little interaction data. Code is available at https://github.com/yanweiw/noise2ptz.

Temporal Logic Imitation: Learning Plan-Satisficing Motion Policies from Demonstrations

Learning from demonstration (LfD) methods have shown promise for solving multi-step tasks; however, these approaches do not guarantee successful reproduction of the task given disturbances. In this work, we identify the roots of such a challenge as the failure of the learned continuous policy to satisfy the discrete plan implicit in the demonstration. By utilizing modes (rather than subgoals) as the discrete abstraction and motion policies with both mode invariance and goal reachability properties, we prove our learned continuous policy can simulate any discrete plan specified by a Linear Temporal Logic (LTL) formula. Consequently, the imitator is robust to both task- and motion-level disturbances and guaranteed to achieve task success. Project page: https://sites.google.com/view/ltl-ds