Ram Ramrakhya

Ram Ramrakhya, Eric Undersander, Dhruv Batra et al. · meta-ai

We present a large-scale study of imitating human demonstrations on tasks that require a virtual robot to search for objects in new environments -- (1) ObjectGoal Navigation (e.g. 'find & go to a chair') and (2) Pick&Place (e.g. 'find mug, pick mug, find counter, place mug on counter'). First, we develop a virtual teleoperation data-collection infrastructure -- connecting Habitat simulator running in a web browser to Amazon Mechanical Turk, allowing remote users to teleoperate virtual robots, safely and at scale. We collect 80k demonstrations for ObjectNav and 12k demonstrations for Pick&Place, which is an order of magnitude larger than existing human demonstration datasets in simulation or on real robots. Second, we attempt to answer the question -- how does large-scale imitation learning (IL) (which hasn't been hitherto possible) compare to reinforcement learning (RL) (which is the status quo)? On ObjectNav, we find that IL (with no bells or whistles) using 70k human demonstrations outperforms RL using 240k agent-gathered trajectories. The IL-trained agent demonstrates efficient object-search behavior -- it peeks into rooms, checks corners for small objects, turns in place to get a panoramic view -- none of these are exhibited as prominently by the RL agent, and to induce these behaviors via RL would require tedious reward engineering. Finally, accuracy vs. training data size plots show promising scaling behavior, suggesting that simply collecting more demonstrations is likely to advance the state of art further. On Pick&Place, the comparison is starker -- IL agents achieve ${\sim}$18% success on episodes with new object-receptacle locations when trained with 9.5k human demonstrations, while RL agents fail to get beyond 0%. Overall, our work provides compelling evidence for investing in large-scale imitation learning. Project page: https://ram81.github.io/projects/habitat-web.

Ram Ramrakhya, Dhruv Batra, Erik Wijmans et al. · mit

We study ObjectGoal Navigation -- where a virtual robot situated in a new environment is asked to navigate to an object. Prior work has shown that imitation learning (IL) using behavior cloning (BC) on a dataset of human demonstrations achieves promising results. However, this has limitations -- 1) BC policies generalize poorly to new states, since the training mimics actions not their consequences, and 2) collecting demonstrations is expensive. On the other hand, reinforcement learning (RL) is trivially scalable, but requires careful reward engineering to achieve desirable behavior. We present PIRLNav, a two-stage learning scheme for BC pretraining on human demonstrations followed by RL-finetuning. This leads to a policy that achieves a success rate of $65.0\%$ on ObjectNav ($+5.0\%$ absolute over previous state-of-the-art). Using this BC$\rightarrow$RL training recipe, we present a rigorous empirical analysis of design choices. First, we investigate whether human demonstrations can be replaced with `free' (automatically generated) sources of demonstrations, e.g. shortest paths (SP) or task-agnostic frontier exploration (FE) trajectories. We find that BC$\rightarrow$RL on human demonstrations outperforms BC$\rightarrow$RL on SP and FE trajectories, even when controlled for same BC-pretraining success on train, and even on a subset of val episodes where BC-pretraining success favors the SP or FE policies. Next, we study how RL-finetuning performance scales with the size of the BC pretraining dataset. We find that as we increase the size of BC-pretraining dataset and get to high BC accuracies, improvements from RL-finetuning are smaller, and that $90\%$ of the performance of our best BC$\rightarrow$RL policy can be achieved with less than half the number of BC demonstrations. Finally, we analyze failure modes of our ObjectNav policies, and present guidelines for further improving them.

Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar et al. · meta-ai

How should we learn visual representations for embodied agents that must see and move? The status quo is tabula rasa in vivo, i.e. learning visual representations from scratch while also learning to move, potentially augmented with auxiliary tasks (e.g. predicting the action taken between two successive observations). In this paper, we show that an alternative 2-stage strategy is far more effective: (1) offline pretraining of visual representations with self-supervised learning (SSL) using large-scale pre-rendered images of indoor environments (Omnidata), and (2) online finetuning of visuomotor representations on specific tasks with image augmentations under long learning schedules. We call this method Offline Visual Representation Learning (OVRL). We conduct large-scale experiments - on 3 different 3D datasets (Gibson, HM3D, MP3D), 2 tasks (ImageNav, ObjectNav), and 2 policy learning algorithms (RL, IL) - and find that the OVRL representations lead to significant across-the-board improvements in state of art, on ImageNav from 29.2% to 54.2% (+25% absolute, 86% relative) and on ObjectNav from 18.1% to 23.2% (+5.1% absolute, 28% relative). Importantly, both results were achieved by the same visual encoder generalizing to datasets that were not seen during pretraining. While the benefits of pretraining sometimes diminish (or entirely disappear) with long finetuning schedules, we find that OVRL's performance gains continue to increase (not decrease) as the agent is trained for 2 billion frames of experience.

Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya et al. · meta-ai

We present a single neural network architecture composed of task-agnostic components (ViTs, convolutions, and LSTMs) that achieves state-of-art results on both the ImageNav ("go to location in <this picture>") and ObjectNav ("find a chair") tasks without any task-specific modules like object detection, segmentation, mapping, or planning modules. Such general-purpose methods offer advantages of simplicity in design, positive scaling with available compute, and versatile applicability to multiple tasks. Our work builds upon the recent success of self-supervised learning (SSL) for pre-training vision transformers (ViT). However, while the training recipes for convolutional networks are mature and robust, the recipes for ViTs are contingent and brittle, and in the case of ViTs for visual navigation, yet to be fully discovered. Specifically, we find that vanilla ViTs do not outperform ResNets on visual navigation. We propose the use of a compression layer operating over ViT patch representations to preserve spatial information along with policy training improvements. These improvements allow us to demonstrate positive scaling laws for the first time in visual navigation tasks. Consequently, our model advances state-of-the-art performance on ImageNav from 54.2% to 82.0% success and performs competitively against concurrent state-of-art on ObjectNav with success rate of 64.0% vs. 65.0%. Overall, this work does not present a fundamentally new approach, but rather recommendations for training a general-purpose architecture that achieves state-of-art performance today and could serve as a strong baseline for future methods.

Sriram Yenamandra, Arun Ramachandran, Mukul Khanna et al. · cmu

In order to develop robots that can effectively serve as versatile and capable home assistants, it is crucial for them to reliably perceive and interact with a wide variety of objects across diverse environments. To this end, we proposed Open Vocabulary Mobile Manipulation as a key benchmark task for robotics: finding any object in a novel environment and placing it on any receptacle surface within that environment. We organized a NeurIPS 2023 competition featuring both simulation and real-world components to evaluate solutions to this task. Our baselines on the most challenging version of this task, using real perception in simulation, achieved only an 0.8% success rate; by the end of the competition, the best participants achieved an 10.8\% success rate, a 13x improvement. We observed that the most successful teams employed a variety of methods, yet two common threads emerged among the best solutions: enhancing error detection and recovery, and improving the integration of perception with decision-making processes. In this paper, we detail the results and methodologies used, both in simulation and real-world settings. We discuss the lessons learned and their implications for future research. Additionally, we compare performance in real and simulated environments, emphasizing the necessity for robust generalization to novel settings.

Alexander H. Liu, Alexis Tacnet, Andy Ehrenberg et al. · deepmind, tsinghua

We introduce Voxtral TTS, an expressive multilingual text-to-speech model that generates natural speech from as little as 3 seconds of reference audio. Voxtral TTS adopts a hybrid architecture that combines auto-regressive generation of semantic speech tokens with flow-matching for acoustic tokens. These tokens are encoded and decoded with Voxtral Codec, a speech tokenizer trained from scratch with a hybrid VQ-FSQ quantization scheme. In human evaluations conducted by native speakers, Voxtral TTS is preferred for multilingual voice cloning due to its naturalness and expressivity, achieving a 68.4\% win rate over ElevenLabs Flash v2.5. We release the model weights under a CC BY-NC license.

Karmesh Yadav, Ram Ramrakhya, Santhosh Kumar Ramakrishnan et al.

We present the Habitat-Matterport 3D Semantics (HM3DSEM) dataset. HM3DSEM is the largest dataset of 3D real-world spaces with densely annotated semantics that is currently available to the academic community. It consists of 142,646 object instance annotations across 216 3D spaces and 3,100 rooms within those spaces. The scale, quality, and diversity of object annotations far exceed those of prior datasets. A key difference setting apart HM3DSEM from other datasets is the use of texture information to annotate pixel-accurate object boundaries. We demonstrate the effectiveness of HM3DSEM dataset for the Object Goal Navigation task using different methods. Policies trained using HM3DSEM perform outperform those trained on prior datasets. Introduction of HM3DSEM in the Habitat ObjectNav Challenge lead to an increase in participation from 400 submissions in 2021 to 1022 submissions in 2022.

Naoki Yokoyama, Ram Ramrakhya, Abhishek Das et al.

We present the Habitat-Matterport 3D Open Vocabulary Object Goal Navigation dataset (HM3D-OVON), a large-scale benchmark that broadens the scope and semantic range of prior Object Goal Navigation (ObjectNav) benchmarks. Leveraging the HM3DSem dataset, HM3D-OVON incorporates over 15k annotated instances of household objects across 379 distinct categories, derived from photo-realistic 3D scans of real-world environments. In contrast to earlier ObjectNav datasets, which limit goal objects to a predefined set of 6-20 categories, HM3D-OVON facilitates the training and evaluation of models with an open-set of goals defined through free-form language at test-time. Through this open-vocabulary formulation, HM3D-OVON encourages progress towards learning visuo-semantic navigation behaviors that are capable of searching for any object specified by text in an open-vocabulary manner. Additionally, we systematically evaluate and compare several different types of approaches on HM3D-OVON. We find that HM3D-OVON can be used to train an open-vocabulary ObjectNav agent that achieves both higher performance and is more robust to localization and actuation noise than the state-of-the-art ObjectNav approach. We hope that our benchmark and baseline results will drive interest in developing embodied agents that can navigate real-world spaces to find household objects specified through free-form language, taking a step towards more flexible and human-like semantic visual navigation. Code and videos available at: naoki.io/ovon.

Yuhao Yang, Zhen Yang, Zi-Yi Dou et al.

Multimodal agents for computer use rely exclusively on primitive actions (click, type, scroll) that require accurate visual grounding and lengthy execution chains, leading to cascading failures and performance bottlenecks. While other agents leverage rich programmatic interfaces (APIs, MCP servers, tools), computer-use agents (CUAs) remain isolated from these capabilities. We present UltraCUA, a foundation model that bridges this gap through hybrid action -- seamlessly integrating GUI primitives with high-level programmatic tool calls. To achieve this, our approach comprises four key components: (1) an automated pipeline that scales programmatic tools from software documentation, open-source repositories, and code generation; (2) a synthetic data engine producing over 17,000 verifiable tasks spanning real-world computer-use scenarios; (3) a large-scale high-quality hybrid action trajectory collection with both low-level GUI actions and high-level programmatic tool calls; and (4) a two-stage training pipeline combining supervised fine-tuning with online reinforcement learning, enabling strategic alternation between low-level and high-level actions. Experiments with our 7B and 32B models demonstrate substantial improvements over state-of-the-art agents. On OSWorld, UltraCUA models achieve an average 22% relative improvement over base models, while being 11% faster in terms of steps. Out-of-domain evaluation on WindowsAgentArena shows our model reaches 21.7% success rate, outperforming baselines trained on Windows data. The hybrid action mechanism proves critical, reducing error propagation while maintaining execution efficiency.

Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani et al.

The Embodied AI community has made significant strides in visual navigation tasks, exploring targets from 3D coordinates, objects, language descriptions, and images. However, these navigation models often handle only a single input modality as the target. With the progress achieved so far, it is time to move towards universal navigation models capable of handling various goal types, enabling more effective user interaction with robots. To facilitate this goal, we propose GOAT-Bench, a benchmark for the universal navigation task referred to as GO to AnyThing (GOAT). In this task, the agent is directed to navigate to a sequence of targets specified by the category name, language description, or image in an open-vocabulary fashion. We benchmark monolithic RL and modular methods on the GOAT task, analyzing their performance across modalities, the role of explicit and implicit scene memories, their robustness to noise in goal specifications, and the impact of memory in lifelong scenarios.

Matthew Chang, Gunjan Chhablani, Alexander Clegg et al.

We present a benchmark for Planning And Reasoning Tasks in humaN-Robot collaboration (PARTNR) designed to study human-robot coordination in household activities. PARTNR tasks exhibit characteristics of everyday tasks, such as spatial, temporal, and heterogeneous agent capability constraints. We employ a semi-automated task generation pipeline using Large Language Models (LLMs), incorporating simulation in the loop for grounding and verification. PARTNR stands as the largest benchmark of its kind, comprising 100,000 natural language tasks, spanning 60 houses and 5,819 unique objects. We analyze state-of-the-art LLMs on PARTNR tasks, across the axes of planning, perception and skill execution. The analysis reveals significant limitations in SoTA models, such as poor coordination and failures in task tracking and recovery from errors. When LLMs are paired with real humans, they require 1.5x as many steps as two humans collaborating and 1.1x more steps than a single human, underscoring the potential for improvement in these models. We further show that fine-tuning smaller LLMs with planning data can achieve performance on par with models 9 times larger, while being 8.6x faster at inference. Overall, PARTNR highlights significant challenges facing collaborative embodied agents and aims to drive research in this direction.

Ram Ramrakhya, Aniruddha Kembhavi, Dhruv Batra et al. · allen-ai

Computer vision tasks typically involve describing what is present in an image (e.g. classification, detection, segmentation, and captioning). We study a visual common sense task that requires understanding what is not present. Specifically, given an image (e.g. of a living room) and name of an object ("cushion"), a vision system is asked to predict semantically-meaningful regions (masks or bounding boxes) in the image where that object could be placed or is likely be placed by humans (e.g. on the sofa). We call this task: Semantic Placement (SP) and believe that such common-sense visual understanding is critical for assitive robots (tidying a house), and AR devices (automatically rendering an object in the user's space). Studying the invisible is hard. Datasets for image description are typically constructed by curating relevant images and asking humans to annotate the contents of the image; neither of those two steps are straightforward for objects not present in the image. We overcome this challenge by operating in the opposite direction: we start with an image of an object in context from web, and then remove that object from the image via inpainting. This automated pipeline converts unstructured web data into a dataset comprising pairs of images with/without the object. Using this, we collect a novel dataset, with ${\sim}1.3$M images across $9$ object categories, and train a SP prediction model called CLIP-UNet. CLIP-UNet outperforms existing VLMs and baselines that combine semantic priors with object detectors on real-world and simulated images. In our user studies, we find that the SP masks predicted by CLIP-UNet are favored $43.7\%$ and $31.3\%$ times when comparing against the $4$ SP baselines on real and simulated images. In addition, we demonstrate leveraging SP mask predictions from CLIP-UNet enables downstream applications like building tidying robots in indoor environments.

Ram Ramrakhya, Matthew Chang, Xavier Puig et al.

Embodied agents operating in household environments must interpret ambiguous and under-specified human instructions. A capable household robot should recognize ambiguity and ask relevant clarification questions to infer the user intent accurately, leading to more effective task execution. To study this problem, we introduce the Ask-to-Act task, where an embodied agent is tasked with a single or multi-object rearrangement task using an under-specified instruction in a home environment. The agent must strategically ask minimal, yet relevant, clarification questions to resolve ambiguity while navigating under partial observability. To address this challenge, we propose a novel approach that fine-tunes multi-modal large language models (MLLMs) as vision-language-action (VLA) policies using online reinforcement learning (RL) with LLM-generated rewards. Our method eliminates the need for large-scale human demonstrations or manually engineered rewards for training such agents. We benchmark against strong zero-shot baselines including GPT-4o as well as supervised fine-tuned MLLMs on our task. Our results show that our RL-finetuned MLLM outperforms all baselines by a significant margin (10.4-16.5%), generalizing well to novel scenes and tasks. To the best of our knowledge, this is the first demonstration of adapting MLLMs as VLA agents that can act and ask for help using LLM-generated rewards with online RL.

Ram Ramrakhya, Andrew Szot, Omar Attia et al. · apple-ml, gatech

Post-Training Multimodal Large Language Models (MLLMs) to build interactive agents holds promise across domains such as computer-use, web navigation, and robotics. A key challenge in scaling such post-training is lack of high-quality downstream agentic task datasets with tasks that are diverse, feasible, and verifiable. Existing approaches for task generation rely heavily on human annotation or prompting MLLM with limited downstream environment information, which is either costly or poorly scalable as it yield tasks with limited coverage. To remedy this, we present AutoPlay, a scalable pipeline for task generation that explicitly explores interactive environments to discover possible interactions and current state information to synthesize environment-grounded tasks. AutoPlay operates in two stages: (i) an exploration phase, where an MLLM explorer agent systematically uncovers novel environment states and functionalities, and (ii) a task generation phase, where a task generator leverages exploration trajectories and a set of task guideline prompts as context to synthesize diverse, executable, and verifiable tasks. We show AutoPlay generates 20k tasks across 20 Android applications and 10k tasks across 13 applications Ubuntu applications to train mobile-use and computer-use agents. AutoPlay generated tasks enable large-scale task demonstration synthesis without human annotation by employing an MLLM task executor and verifier. This data enables training MLLM-based UI agents that improve success rates up to $20.0\%$ on mobile-use and $10.9\%$ on computer-use scenarios. In addition, AutoPlay generated tasks combined with MLLM verifier-based rewards enable scaling reinforcement learning training of UI agents, leading to an additional $5.7\%$ gain. coverage. These results establish AutoPlay as a scalable approach for post-training capable MLLM agents reducing reliance on human annotation.

Zhen Yang, Zi-Yi Dou, Di Feng et al.

Developing autonomous agents that effectively interact with Graphic User Interfaces (GUIs) remains a challenging open problem, especially for small on-device models. In this paper, we present Ferret-UI Lite, a compact, end-to-end GUI agent that operates across diverse platforms, including mobile, web, and desktop. Utilizing techniques optimized for developing small models, we build our 3B Ferret-UI Lite agent through curating a diverse GUI data mixture from real and synthetic sources, strengthening inference-time performance through chain-of-thought reasoning and visual tool-use, and reinforcement learning with designed rewards. Ferret-UI Lite achieves competitive performance with other small-scale GUI agents. In GUI grounding, Ferret-UI Lite attains scores of $91.6\%$, $53.3\%$, and $61.2\%$ on the ScreenSpot-V2, ScreenSpot-Pro, and OSWorld-G benchmarks, respectively. For GUI navigation, Ferret-UI Lite achieves success rates of $28.0\%$ on AndroidWorld and $19.8\%$ on OSWorld. We share our methods and lessons learned from developing compact, on-device GUI agents.

Alexander H. Liu, Andy Ehrenberg, Andy Lo et al.

We introduce Voxtral Realtime, a natively streaming automatic speech recognition model that matches offline transcription quality at sub-second latency. Unlike approaches that adapt offline models through chunking or sliding windows, Voxtral Realtime is trained end-to-end for streaming, with explicit alignment between audio and text streams. Our architecture builds on the Delayed Streams Modeling framework, introducing a new causal audio encoder and Ada RMS-Norm for improved delay conditioning. We scale pretraining to a large-scale dataset spanning 13 languages. At a delay of 480ms, Voxtral Realtime achieves performance on par with Whisper, the most widely deployed offline transcription system. We release the model weights under the Apache 2.0 license.

Utsav Garg, Viraj Prabhu, Deshraj Yadav et al.

We present Fabrik, an online neural network editor that provides tools to visualize, edit, and share neural networks from within a browser. Fabrik provides a simple and intuitive GUI to import neural networks written in popular deep learning frameworks such as Caffe, Keras, and TensorFlow, and allows users to interact with, build, and edit models via simple drag and drop. Fabrik is designed to be framework agnostic and support high interoperability, and can be used to export models back to any supported framework. Finally, it provides powerful collaborative features to enable users to iterate over model design remotely and at scale.