Letian Fu

Ilija Radosavovic, Baifeng Shi, Letian Fu et al.

We present a self-supervised sensorimotor pre-training approach for robotics. Our model, called RPT, is a Transformer that operates on sequences of sensorimotor tokens. Given a sequence of camera images, proprioceptive robot states, and actions, we encode the sequence into tokens, mask out a subset, and train a model to predict the missing content from the rest. We hypothesize that if a robot can predict the masked-out content it will have acquired a good model of the physical world that can enable it to act. RPT is designed to operate on latent visual representations which makes prediction tractable, enables scaling to larger models, and allows fast inference on a real robot. To evaluate our approach, we collected a dataset of 20,000 real-world trajectories over 9 months using a combination of motion planning and grasping algorithms. We find that sensorimotor pre-training consistently outperforms training from scratch, has favorable scaling properties, and enables transfer across different tasks, environments, and robots.

Letian Fu, Huang Huang, Gaurav Datta et al.

We explore how to enhance next-token prediction models to perform in-context imitation learning on a real robot, where the robot executes new tasks by interpreting contextual information provided during the input phase, without updating its underlying policy parameters. We propose In-Context Robot Transformer (ICRT), a causal transformer that performs autoregressive prediction on sensorimotor trajectories without relying on any linguistic data or reward function. This formulation enables flexible and training-free execution of new tasks at test time, achieved by prompting the model with sensorimotor trajectories of the new task composing of image observations, actions and states tuples, collected through human teleoperation. Experiments with a Franka Emika robot demonstrate that the ICRT can adapt to new tasks specified by prompts, even in environment configurations that differ from both the prompt and the training data. In a multitask environment setup, ICRT significantly outperforms current state-of-the-art next-token prediction models in robotics on generalizing to unseen tasks. Code, checkpoints and data are available on https://icrt.dev/

Andrew Goldberg, Kavish Kondap, Tianshuang Qiu et al.

Generative AI systems have shown impressive capabilities in creating text, code, and images. Inspired by the rich history of research in industrial ''Design for Assembly'', we introduce a novel problem: Generative Design-for-Robot-Assembly (GDfRA). The task is to generate an assembly based on a natural language prompt (e.g., ''giraffe'') and an image of available physical components, such as 3D-printed blocks. The output is an assembly, a spatial arrangement of these components, and instructions for a robot to build this assembly. The output must 1) resemble the requested object and 2) be reliably assembled by a 6 DoF robot arm with a suction gripper. We then present Blox-Net, a GDfRA system that combines generative vision language models with well-established methods in computer vision, simulation, perturbation analysis, motion planning, and physical robot experimentation to solve a class of GDfRA problems with minimal human supervision. Blox-Net achieved a Top-1 accuracy of 63.5% in the ''recognizability'' of its designed assemblies (eg, resembling giraffe as judged by a VLM). These designs, after automated perturbation redesign, were reliably assembled by a robot, achieving near-perfect success across 10 consecutive assembly iterations with human intervention only during reset prior to assembly. Surprisingly, this entire design process from textual word (''giraffe'') to reliable physical assembly is performed with zero human intervention.

Letian Fu, Gaurav Datta, Huang Huang et al.

Touch is an important sensing modality for humans, but it has not yet been incorporated into a multimodal generative language model. This is partially due to the difficulty of obtaining natural language labels for tactile data and the complexity of aligning tactile readings with both visual observations and language descriptions. As a step towards bridging that gap, this work introduces a new dataset of 44K in-the-wild vision-touch pairs, with English language labels annotated by humans (10%) and textual pseudo-labels from GPT-4V (90%). We use this dataset to train a vision-language-aligned tactile encoder for open-vocabulary classification and a touch-vision-language (TVL) model for text generation using the trained encoder. Results suggest that by incorporating touch, the TVL model improves (+29% classification accuracy) touch-vision-language alignment over existing models trained on any pair of those modalities. Although only a small fraction of the dataset is human-labeled, the TVL model demonstrates improved visual-tactile understanding over GPT-4V (+12%) and open-source vision-language models (+32%) on a new touch-vision understanding benchmark. Code and data: https://tactile-vlm.github.io.

Zirui Wang, Junyi Zhang, Jiaxin Ge et al.

Modern Vision-Language Models (VLMs) remain poorly characterized in multi-step visual interactions, particularly in how they integrate perception, memory, and action over long horizons. We introduce VisGym, a gymnasium of 17 environments for evaluating and training VLMs. The suite spans symbolic puzzles, real-image understanding, navigation, and manipulation, and provides flexible controls over difficulty, input representation, planning horizon, and feedback. We also provide multi-step solvers that generate structured demonstrations, enabling supervised finetuning. Our evaluations show that all frontier models struggle in interactive settings, achieving low success rates in both the easy (46.6%) and hard (26.0%) configurations. Our experiments reveal notable limitations: models struggle to effectively leverage long context, performing worse with an unbounded history than with truncated windows. Furthermore, we find that several text-based symbolic tasks become substantially harder once rendered visually. However, explicit goal observations, textual feedback, and exploratory demonstrations in partially observable or unknown-dynamics settings for supervised finetuning yield consistent gains, highlighting concrete failure modes and pathways for improving multi-step visual decision-making. Code, data, and models can be found at: https://visgym.github.io/.

Kaiyuan Chen, Letian Fu, David Huang et al.

Recent results suggest that very large datasets of teleoperated robot demonstrations can be used to train transformer-based models that have the potential to generalize to new scenes, robots, and tasks. However, curating, distributing, and loading large datasets of robot trajectories, which typically consist of video, textual, and numerical modalities - including streams from multiple cameras - remains challenging. We propose Robo-DM, an efficient open-source cloud-based data management toolkit for collecting, sharing, and learning with robot data. With Robo-DM, robot datasets are stored in a self-contained format with Extensible Binary Meta Language (EBML). Robo-DM can significantly reduce the size of robot trajectory data, transfer costs, and data load time during training. Compared to the RLDS format used in OXE datasets, Robo-DM's compression saves space by up to 70x (lossy) and 3.5x (lossless). Robo-DM also accelerates data retrieval by load-balancing video decoding with memory-mapped decoding caches. Compared to LeRobot, a framework that also uses lossy video compression, Robo-DM is up to 50x faster when decoding sequentially. We physically evaluate a model trained by Robo-DM with lossy compression, a pick-and-place task, and In-Context Robot Transformer. Robo-DM uses 75x compression of the original dataset and does not suffer reduction in downstream task accuracy.

Huang Huang, Fangchen Liu, Letian Fu et al.

Vision-Language-Action (VLA) models aim to predict robotic actions based on visual observations and language instructions. Existing approaches require fine-tuning pre-trained visionlanguage models (VLMs) as visual and language features are independently fed into downstream policies, degrading the pre-trained semantic alignments. We propose OTTER, a novel VLA architecture that leverages these existing alignments through explicit, text-aware visual feature extraction. Instead of processing all visual features, OTTER selectively extracts and passes only task-relevant visual features that are semantically aligned with the language instruction to the policy transformer. This allows OTTER to keep the pre-trained vision-language encoders frozen. Thereby, OTTER preserves and utilizes the rich semantic understanding learned from large-scale pre-training, enabling strong zero-shot generalization capabilities. In simulation and real-world experiments, OTTER significantly outperforms existing VLA models, demonstrating strong zeroshot generalization to novel objects and environments. Video, code, checkpoints, and dataset: https://ottervla.github.io/.

Letian Fu, Long Lian, Renhao Wang et al.

In this work, we examine the impact of inter-patch dependencies in the decoder of masked autoencoders (MAE) on representation learning. We decompose the decoding mechanism for masked reconstruction into self-attention between mask tokens and cross-attention between masked and visible tokens. Our findings reveal that MAE reconstructs coherent images from visible patches not through interactions between patches in the decoder but by learning a global representation within the encoder. This discovery leads us to propose a simple visual pretraining framework: cross-attention masked autoencoders (CrossMAE). This framework employs only cross-attention in the decoder to independently read out reconstructions for a small subset of masked patches from encoder outputs. This approach achieves comparable or superior performance to traditional MAE across models ranging from ViT-S to ViT-H and significantly reduces computational requirements. By its design, CrossMAE challenges the necessity of interaction between mask tokens for effective masked pretraining. Code and models are publicly available: https://crossmae.github.io

Huang Huang, Michael Danielczuk, Chung Min Kim et al.

Shelves are common in homes, warehouses, and commercial settings due to their storage efficiency. However, this efficiency comes at the cost of reduced visibility and accessibility. When looking from a side (lateral) view of a shelf, most objects will be fully occluded, resulting in a constrained lateral-access mechanical search problem. To address this problem, we introduce: (1) a novel bluction tool, which combines a thin pushing blade and suction cup gripper, (2) an improved LAX-RAY simulation pipeline and perception model that combines ray-casting with 2D Minkowski sums to efficiently generate target occupancy distributions, and (3) a novel SLAX-RAY search policy, which optimally reduces target object distribution support area using the bluction tool. Experimental data from 2000 simulated shelf trials and 18 trials with a physical Fetch robot equipped with the bluction tool suggest that using suction grasping actions improves the success rate over the highest performing push-only policy by 26% in simulation and 67% in physical environments.

Letian Fu, Michael Danielczuk, Ashwin Balakrishna et al.

While deep learning has enabled significant progress in designing general purpose robot grasping systems, there remain objects which still pose challenges for these systems. Recent work on Exploratory Grasping has formalized the problem of systematically exploring grasps on these adversarial objects and explored a multi-armed bandit model for identifying high-quality grasps on each object stable pose. However, these systems are still limited to exploring a small number or grasps on each object. We present Learned Efficient Grasp Sets (LEGS), an algorithm that efficiently explores thousands of possible grasps by maintaining small active sets of promising grasps and determining when it can stop exploring the object with high confidence. Experiments suggest that LEGS can identify a high-quality grasp more efficiently than prior algorithms which do not use active sets. In simulation experiments, we measure the gap between the success probability of the best grasp identified by LEGS, baselines, and the most-robust grasp (verified ground truth). After 3000 exploration steps, LEGS outperforms baseline algorithms on 10/14 and 25/39 objects on the Dex-Net Adversarial and EGAD! datasets respectively. We then evaluate LEGS in physical experiments; trials on 3 challenging objects suggest that LEGS converges to high-performing grasps significantly faster than baselines. See https://sites.google.com/view/legs-exp-grasping for supplemental material and videos.