Hi! I am a robotics researcher at NVIDIA on the Isaac Loco-Manipulation team, working on RL post-training of robot foundation models. I completed my PhD at the University of Southern California, where I worked with Jesse Thomason on building and evaluating language-guided robots.

I have previously worked on mobility foundation models at NVIDIA, assistive robotics for mobility limitations at Cornell with Tapo Bhattacharjee, and human-robot interaction at UT Austin with Justin Hart and Peter Stone. I also spent time at Sandia National Labs with James Aimone and Craig Vineyard.

News

Research

2026

Anticipating Unintended Robot Behaviors with Side Effect Critics

Ryan Lindeborg*, Richard Peng*, Ayush Jain, Jesse Zhang, Abrar Anwar, Jesse Thomason

We introduce the concept of Side Effect Critics to anticipate and mitigate unintended robot behaviors when deploying learned policies in real-world environments.

Workshop on Rethinking What It Means to be "Safe" for Generalist Robots, 2026 (Spotlight)

paper

Mechanistic Finetuning of Vision-Language-Action Models via Few-Shot Demonstrations

Chancharik Mitra*, Yusen Luo*, Raj Saravanan*, Dantong Niu, Anirudh Pai, Jesse Thomason, Trevor Darrell, Abrar Anwar, Deva Ramanan, Roei Herzig

Robotic Steering is a mechanistic, few-shot finetuning approach for vision-language-action models that selectively adapts task-specific attention heads, achieving more robust, efficient, and interpretable robot learning than LoRA across diverse tasks.

ECCV 2026 (Oral Presentation)

arxiv

SAFECAST: Robust Failure Detection for VLA Policies with Contrast-Set Training and Calibration

Harshitha Rajaprakash, Aditeya Prajapati, Rong Xue, Abrar Anwar, Jesse Thomason

VLA policies fail under deployment-time shifts like clutter, lighting changes, and reworded instructions. SAFECAST uses contrast set perturbations to train and calibrate hidden-state risk probes, improving failure detection ROC-AUC over state-of-the-art baselines on both real-world DROID and LIBERO simulation.

In Submission

arxiv

DiPS: Dialogue Policy Selection for High-Stakes Persuasion Agents

Tianyi Zhang, Mousumi Das, Abrar Anwar, Jesse Thomason, David Traum

DiPS is a Q-learning framework that dynamically selects persuasion strategies based on the evolving conversational context. In a fire-rescue evacuation scenario, DiPS achieves higher persuasion success than zero-shot LLM and RAG-augmented approaches in both simulated and real human interactions.

SIGDIAL 2026 (Best paper award)

paper arxiv

Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons

Anthony Liang*, Yigit Korkmaz*, Jiahui Zhang, Minyoung Hwang, Abrar Anwar, Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Biyik, Jesse Zhang

Robometer is a general-purpose, video-language-input dense reward model trained on RBM-1M, a dataset of over 1M trajectories spanning 21 robot embodiments. It improves robot learning across online RL, offline RL, model-based RL, failure detection, and data retrieval for imitation learning.

RSS 2026
RL4IL Workshop @ ICRA 2026 (Oral Presentation)

arxiv website

2025

Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning

NVIDIA + me!

I helped introduce approaches for VLA post-training with IsaacLab!

Whitepaper

arxiv

RobotFleet: An Open-Source Framework for Centralized Multi-Robot Task Planning

Rohan Gupta*, Trevor Asbery*, Zain Merchant*, Abrar Anwar, Jesse Thomason

RobotFleet is an open-source, extensible framework that introduces a centralized, modular autonomy stack to simplify scalable planning, scheduling, and execution for heterogeneous multi-robot fleets in open-world tasks.

Multi-Robot Systems Workshop @ RSS 2025.

arxiv code

ReWiND: Language-Guided Rewards Teach Robot Policies without New Demonstrations

Jiahui Zhang*, Yusen Luo*, Abrar Anwar*, Sumedh Sontakke, Joseph Lim, Jesse Thomason, Erdem Biyik, Jesse Zhaing

We design ReWiND rewards that use language-guided rewards to train bimanual arms on OOD tasks in 1 hour! We use offline-to-online, lang-conditioned, visual RL on action-chunked transformers on a real robot and in simulation!

CoRL 2025 (Oral Presentation)
RoboReps Workshop @ RSS 2025 (Best paper nominee)
OOD Workshop @ RSS 2025 (Best paper)

arxiv website twitter

Efficient Evaluation of Multi-Task Robot Policies With Active Experiment Selection

Abrar Anwar*, Rohan Gupta, Zain Merchant, Sayan Ghosh, Willie Neiswanger, Jesse Thomason

The space of language commands a robot can execute grows combinatorially with scene complexity. Evaluating a robot on this large domain is impractical + takes time, so we introduce contrast sets for robots to make small perturbations to test instances. This leads to good test set estimation and less experimenter effort.

CoRL 2025
RobotEvaluation Workshop @ RSS 2025

arxiv twitter

ReMEmbR: Building and Reasoning Over Long-Horizon Spatio-Temporal Memory for Robot Navigation

Abrar Anwar*, John Welsh, Joydeep Biswas, Soha Pouya, Yan Chang

To tackle long-horizon spatio-temporal memory, we introduce a retrieval-based approach for building and reasoning over memory. We show improved performance on planning and embodied question answering given a long video history.

ICRA 2025

arxiv code website blog twitter

M3PT: A Transformer for Multimodal, Multi-Party Social Signal Prediction with Person-aware Blockwise Attention

Yiming Tang, Abrar Anwar, Jesse Thomason

M3PT is a causal multimodal transformer that jointly models social signals across multiple participants and modalities, improving prediction of behaviors like speaking and bite timing in multi-party human interactions.

Nonverbal Cues Workshop @ ICRA 2025 (Best paper)

arxiv code

2024

Contrast Sets for Evaluating Language-Guided Robot Policies

Abrar Anwar*, Rohan Gupta*, Jesse Thomason

The space of language commands a robot can execute grows combinatorially with scene complexity. Evaluating a robot on this large domain is impractical + takes time, so we introduce contrast sets for robots to make small perturbations to test instances. This leads to good test set estimation and less experimenter effort.

CoRL 2024

arxiv twitter

Which One? Leveraging Context Between Objects and Multiple Views for Language Grounding

Chancharik Mitra*, Abrar Anwar*, Rodolfo Corona, Dan Klein, Trevor Darrell, Jesse Thomason

We present the MAGiC model which selects an object referent based on language meant to distinguish between two similar objects by reasoning over both objects from multiple vantage points.

NAACL 2024

paper arxiv code

Generating Contextually-Relevant Navigation Instructions for Blind and Low Vision People

Zain Merchant, Abrar Anwar, Emily Wang, Souti Chattopadhyay, Jesse Thomason

Navigating unfamiliar environments presents significant challenges for blind and low-vision (BLV) individuals. We investigate how grounded instruction generation methods can provide contextually-relevant navigational guidance to BLV users.

ROMAN 2024 Late Breaking Report (LBR)
ROMAN 2024 Interactive AI Workshop (Best paper)

arxiv

Feel the Bite: Robot-Assisted Inside-Mouth Bite Transfer using Robust Mouth Perception and Physical Interaction-Aware Control

Rajat Jenamani, Daniel Stabile, Ziang Liu, Abrar Anwar, Katherine Dimitropoulou, Tapomayukh Bhattacharjee

We design a system to feed people with disabilities in their mouth using real-time mouth perception and tactile-informed control.

HRI 2024. (Best paper nominee)

arxiv video website

2023

Exploring Strategies for Efficient VLN Evaluation

Abrar Anwar*, Rohan Gupta*, Elle Szabo, Jesse Thomason

Evaluation in the real world is often time-consuming and expensive, so we propose a targeted contrast set-based evaluation strategy to efficiently evaluate the linguistic and visual capabilities of an end-to-end VLN policy.

Workshop on Language and Robot Learning (LangRob) @ CoRL 2023. (Oral Presentation)

paper

2022

Human-Robot Commensality: Bite Timing Prediction for Robot-Assisted Feeding in Groups

Jan Ondras*, Abrar Anwar*, Tong Wu*, Fanjun Bu, Malte Jung, Jorge Jose Ortiz, Tapomayukh Bhattacharjee

We develop data-driven models to predict when a robot should feed during social dining scenarios. We build a dataset of human-human commensality, develop novel models to learn social dynamics of when to feed, and conduct a human-robot commensality study.

CoRL 2022

paper arxiv code website

2021

Deep Reinforcement Learning for Optimal Refinement of Cross-Sectional Mesh Sequence Finite Elements

Abrar Anwar

Developed the first deep reinforcement learning framework for mesh refinement and refined “good” quality surface reconstructions of cross-sectional contours using soft-actor critic

Honors Thesis, 2021

Watch Where You're Going! Gaze and Head Orientation as Predictors for Social Robot Navigation

Blake Holman, Abrar Anwar, Akash Singh, Mauricio Tec, Justin Hart, Peter Stone

We leverage virtual reality to collect gaze and position data to create a predictive model and a mixed effects model to show gaze orientation precedes other features

ICRA 2021

paper video

2020

Evolving Spiking Circuit Motifs using Weight Agnostic Neural Networks.

Abrar Anwar, Craig Vineyard, William Severa, Srideep Musuvathy, Suma Cardwell

An evolutionary, weight agnostic method is used to generate spiking neural networks used for classification, control, and various other tasks

AAAI-21 Undergraduate Consortium , 2021
Computer Science Research Institute Summer Proceedings, 2020
International Conference on Neuromorphic Systems 2020 (poster)

paper tech report poster

2019

BrainSLAM: Robust autonomous navigation in sensor-deprived contexts

Felix Wang, James B. Aimone, Abrar Anwar, Srideep Musuvathy.

We explore using brain-inspired approaches to navigation and localization in a noisy, data-sparse environment for a hypersonic glide vehicle. Rotation invariant feature representations are used to increase accuracy and reduce map storage

Sandia National Labs Technical Report, 2019

tech report

Talks