Gautam Salhotra

How Does Contextual Fidelity Impact How We Think, Talk, and Act in AI-assisted Engineering Design? I am excited to share our latest paper with Shantanu Vyas, Ph.D. and Suryapavan Cheruku. As generative AI moves deeper into engineering workflows, we ask a simple but critical question: Should AI be a solution generator, or a reflective thinking partner? To explore this question, we design a mixed-initiative conversational framework where LLMs and VLMs support engineers through reflective questioning, explanations, and suggestions rather than just proposing designs. ​ Our findings show that while AI partners do not immediately improve coarse performance metrics, they can reshape how people explore the design space, reason about trade-offs, and reflect on their decisions, leading to more focused exploration, and lower mental demand. But then, how much should the agent know to be able to help us reflect? To learn more, take a look at our paper and share your thoughts! Please read the paper here (open access): https://lnkd.in/gkGhu-kr Please watch the video here: https://lnkd.in/gfj99Qd5

27

2 Comments

How do you model both equilibrium behavior and time-dependent dynamics within the same framework? In Julia for Engineers Part 2, Dr. Ranjan Anantharaman walks through building, simulating, and analyzing steady-state and dynamic systems using #Dyad. Join us on March 18 to see how these approaches provide deeper insight into system performance under changing conditions. https://lnkd.in/euZ6kHjG #JuliaLang #Dyad #JuliaForEngineers #Modeling #SystemSimulation #ModelBasedDesign #ScientificComputing #Engineering

13

Without Deep Science there is no Deep Tech. this is also why good PhD students are multi-skilled - they can use Physics, Numerical Methods, FEM, all with equal aplomb. If I don't understand the physics of an Oleo strut running multiple LS-DYNA or Ansys is of no consequence. then I cannot design a new landing gear on a new passenger aircraft. India will always remain a two or even three track society. it still thinks it needs MBAs to run space startups. There is a built in lack of proclivity to understand the role of physics in an engineering problem.

5

𝗦𝗶𝗺-𝘁𝗼-𝗥𝗲𝗮𝗹 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗟𝗼𝗰𝗼𝗺𝗼𝘁𝗶𝗼𝗻: 𝗪𝗮𝗹𝗸𝗶𝗻𝗴 𝗮𝗻𝗱 𝗕𝗹𝗶𝗻𝗱 𝗦𝘁𝗮𝗶𝗿 𝗖𝗹𝗶𝗺𝗯𝗶𝗻𝗴 𝗼𝗻 𝗮 𝗨𝗻𝗶𝘁𝗿𝗲𝗲 𝗚𝗼𝟮 The sim-to-real gap is one of the most challenging problems in legged locomotion. For my 10-week independent project at Northwestern's MSR program, I trained RL (PPO Actor-Critic) policies for quadruped locomotion in simulation (𝗚𝗲𝗻𝗲𝘀𝗶𝘀) and deployed them on a real robot. I was able to successfully transfer walking and stair climbing policies to hardware. The major challenge in sim-to-real was the very different actuator and contact dynamics between the two settings. 𝗗𝗼𝗺𝗮𝗶𝗻 𝗿𝗮𝗻𝗱𝗼𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻 helped by making reality a subset of the simulation trials. This gets the real robot working, though looking closely you can see differences in behavior on a one-to-one comparison. Adding 𝘀𝗲𝗻𝘀𝗼𝗿 𝗻𝗼𝗶𝘀𝗲 𝗮𝗻𝗱 𝗹𝗮𝘁𝗲𝗻𝗰𝘆 𝗶𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻 improved the policy significantly as it pushes sim closer to reality. Learning 𝗹𝗲𝗴 𝘀𝘁𝗶𝗳𝗳𝗻𝗲𝘀𝘀 as a policy output made the policy robust across different terrain and improved recovery behavior. The stair climbing policy was fine-tuned from the base walking checkpoint, entirely blind with no camera or LiDAR. The stair heightfield curriculum kicks in first, letting the policy focus on the new terrain, with DR ramping up only after it gets comfortable with the stairs. Training done in 𝗚𝗲𝗻𝗲𝘀𝗶𝘀 (built on a fork of the Genesis repo): https://lnkd.in/gxgFkhHf Deployment on 𝗨𝗻𝗶𝘁𝗿𝗲𝗲 𝗚𝗼𝟮 (built on a fork of the Unitree SDK): https://lnkd.in/gbxBmVcs Full write-up: https://lnkd.in/gKbJ5AGk This work builds heavily on two papers: Extreme Parkour with Legged Robots (https://lnkd.in/gTT2KbKc), for the asymmetric actor-critic, privileged observations, and DR strategy Variable Stiffness for Robust Locomotion through RL (https://lnkd.in/gi-c3EEv), for the per-leg adaptive stiffness formulation and DR parameter ranges Special thanks to Jyothi Swaroop Kasina for the idea brainstorming and hands-on help throughout, and to Conor Wood Hayes and Daniel Augustin for their ideas and testing. #Robotics #ReinforcementLearning #SimToReal #Quadruped #Legged #Locomotion #UnitreeGo2

86

22 Comments

When a human operator makes a mistake, it’s often just an anecdote. When an autonomous robot makes one, it’s a permanent log. That asymmetry changes a lot in how industrial robotics must be built, explained, and ultimately sold. Waymo recently published data claiming their autonomous vehicles crash 92% less than human drivers (link in comment). Beyond the number itself, what struck me is the move behind it: using evidence to make autonomous performance legible enough to earn trust. In consumer autonomy, that trust is largely social. In industrial automation, it becomes operational, commercial, and financial. That is where the bar changes. In B2B robotics, it is no longer enough to promise autonomy. You have to make performance measurable, explainable, and defensible in the context where the system actually operates. For a long time, many robotics companies treated observability mostly as an R&D tool. Dashboards were built for engineers: to debug offboard orchestration, investigate incidents, or tune system behavior. But that is no longer enough. Customers are not only asking whether the system is safe. They are asking whether its performance can be demonstrated in conditions that resemble their own operational reality. And proving that a system works is often harder than making it work. Take something as apparently simple as fleet uptime. Producing a credible number is already an engineering problem. You have to separate genuine system failure from the messy reality of warehouse operations: an unexpected physical intrusion, a cancelled order, a blocked aisle, or a worker pressing an emergency stop for reasons that have little to do with the software itself. If you cannot map that causality with enough confidence, you lose control of the narrative. I have seen projects slow down not because the robots were underperforming, but because the available data was not robust enough to counter the subjective feeling on the warehouse floor. The reverse is also true. If you can show, with enough precision and credibility, that your AMRs handle high-value payloads without the damage patterns seen in manual operations, you are helping your customer reduce risk, justify investment, and trust automation at a different level. I saw a version of this years ago at Stanley Robotics. One important benefit was that robots simply did not scratch cars the way human jockeys sometimes do. The robot moves the payload. The offboard software makes the performance legible. This is why I believe our industry is moving from data as a debugging residue to data as a product capability. What used to be seen as supporting material for engineering is becoming central to how autonomy is understood, trusted, and eventually bought.

11

1 Comment

𝗙𝗘𝗔 𝗰𝗼𝗺𝗽𝘂𝘁𝗮𝘁𝗶𝗼𝗻 𝗵𝗮𝘀 𝗲𝘃𝗼𝗹𝘃𝗲𝗱. This is always an interesting topic, so I thought it would be helpful to share some of my thoughts on it here in this post. 1. Processor performance has increased dramatically, consistent with Moore’s Law, for decades now.  Many factors contribute to computational performance for 𝗙𝗶𝗻𝗶𝘁𝗲 𝗘𝗹𝗲𝗺𝗲𝗻𝘁 𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 (𝗙𝗘𝗔) beyond simply CPU speed.  Improvements in computation are also driven by 𝗽𝗮𝗿𝗮𝗹𝗹𝗲𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻, the use of 𝗚𝗣𝗨𝘀, 𝗮𝗻𝗱 𝗱𝗶𝘀𝘁𝗿𝗶𝗯𝘂𝘁𝗲𝗱 𝗰𝗼𝗺𝗽𝘂𝘁𝗶𝗻𝗴, in addition to raw CPU speed. Instead of purchasing expensive computers, an increasingly popular option is to 𝗿𝗲𝗻𝘁 𝗰𝗹𝗼𝘂𝗱 𝗿𝗲𝘀𝗼𝘂𝗿𝗰𝗲𝘀. 2. A better measure of computational demand in FEA is the number of coupled DOF, not simply the number of elements.  𝗠𝗮𝘁𝗿𝗶𝘅 𝘀𝗽𝗮𝗿𝘀𝗶𝘁𝘆, 𝘀𝗼𝗹𝘃𝗲𝗿 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲, 𝗺𝗲𝗺𝗼𝗿𝘆 𝗯𝗮𝗻𝗱𝘄𝗶𝗱𝘁𝗵, 𝗮𝗻𝗱 𝗻𝘂𝗺𝗲𝗿𝗶𝗰𝗮𝗹 𝘀𝘁𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗮𝗹𝗹 𝗽𝗹𝗮𝘆 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝗿𝗼𝗹𝗲𝘀.  3. Similar to today, since the inception of FEM, engineers have sought accuracy through the use of appropriate solution techniques. While mesh refinement and higher DOF have always been central, modern strategies now include adaptive meshing, higher-order elements, meshless methods for special situations, isogeometric element methods for problems requiring high-order continuity, and sparse iterative solvers that can deliver accuracy with fewer resources. 𝗥𝗲𝗱𝘂𝗰𝗲𝗱-𝗼𝗿𝗱𝗲𝗿 𝘀𝘂𝗿𝗿𝗼𝗴𝗮𝘁𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 have regained popularity due to the advent of machine learning and AI.     4. With increased computational resources and improved FEA methods, over the years, what was once thought to be too costly is now feasible, allowing routine complex analysis, such as nonlinear and multi-physics simulations, to be possible with modest computing resources by today's standards.     5. 𝗗𝗶𝗿𝗲𝗰𝘁 𝗲𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹 𝘃𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻 𝗶𝘀 𝗻𝗼𝘁 𝗮𝗹𝘄𝗮𝘆𝘀 𝗳𝗲𝗮𝘀𝗶𝗯𝗹𝗲. Confidence often stems from a combination of verification (checking the correctness of the model), validation (calibrating material models and comparing them with experimental or benchmark data), and 𝘀𝗲𝗻𝘀𝗶𝘁𝗶𝘃𝗶𝘁𝘆 𝘀𝘁𝘂𝗱𝗶𝗲𝘀.     6. The availability of 𝗳𝗿𝗲𝗲 𝗮𝗰𝗰𝗲𝘀𝘀 to verified finite element models (for example, on Github), both closed (e.g. Ansys, Abaqus) and open-source (e.g., Code_Aster, CalculiX, FEniCS), that can be modified for use in verification and validation of similar models for analysis, can help enable engineers and practitioners to build on existing verified models, accelerating both learning and collaboration. What are your thoughts on how FEA computation has evolved

76

40 Comments

AI for Science is shifting the paradigm in Computational Fluid Dynamics (CFD)—moving us from hours of simulation to seconds of inference. The Solution Architect and PhysicsNeMo team (Abouzar Ghasemi Ira Shokar Farah Hariri Mohammad Amin Nabian Ram Cherukuri) at NVIDIA have done an awesome in developing a learning kit to bridge the gap between AI and CFD. If you are looking to get hands-on with AI Physics, here is a complete walkthrough using the canonical Ahmed Body benchmark. These two Jupyter notebooks guide you through the lifecycle of a deep learning surrogate model using the PhysicsNeMo DoMINO (Decomposable Multi-scale Iterative Neural Operator) architecture. Here is the breakdown: Notebook 1: Data Preprocessing Raw simulation data needs to be ML-ready. We cover the extraction and transformation of the Ahmed Body Surface dataset, tackling geometry handling, point cloud generation, and preparing inputs for the model.  Explore it here: https://lnkd.in/epM_TZGJ Notebook 2: Training the Model We dive straight into configuring and training the DoMINO architecture—specifically designed to handle the multi-scale challenges of 3D aerodynamic flows. You will learn how to set up the training loop, handle loss functions, and validate performance on unseen geometries.  Explore it here: https://lnkd.in/etMTjpWq This workflow is a prime example of End-to-End AI for Science, bridging the gap between traditional CAE data and modern neural operators. Explore the full repository:  https://lnkd.in/eG_div-D

334

3 Comments

Excited to be working with the team at Mind Robotics as they build toward real-world factory deployments! The next generation of robotics will be defined by teams willing to iterate in unforgiving real-world environments, solving the hard problems of operationalizing new models and hardware. They’re building for real tasks, in real factories, at real scale. Letting the factory shape the roadmap.

62

8 Comments

Been following NVIDIA GTC this week, and one thing is pretty clear — GPU-based simulation is really hitting production now. A few examples that stood out: - Honda reporting ~34× speedup for aero sims on Grace Blackwell - JLR and Mercedes-Benz actually deploying GPU CFD workflows - Ascendance running same-day VTOL aerodynamic campaigns (which just wasn’t realistic on CPU clusters) So yeah, simulation is getting a lot faster. But something else Jensen mentioned stuck with me more. He drew a line between: - traditional physics solvers - AI-based physics models (surrogates) Those 34× gains are still the same solvers — just running on better hardware. Surrogates are a different direction. They don’t solve the equations every time. They learn from simulation data and then predict almost instantly. And the interesting part is how these two things connect. The push toward GPU solvers isn’t just about speed — it’s also about generating way more data, much faster. And that data is exactly what these AI models need. So in a way, what’s happening right now is: faster simulation → more data → better surrogates NVIDIA seems to be leaning into both sides of this. Faster solvers, and also things like Apollo (their physics AI model stack). Feels like we’re moving from: “how fast can we simulate?” to “when do we even need to simulate?” Curious how others are thinking about this, especially people working in CFD / HPC / SciML.

31

1 Comment

Thrilled to be co-organizing the Symposium on Robotics & Automation in Smart Manufacturing Systems with @Haseung Chung, Amit Bagchi, Guangze Li, and Gary Zhang at ASME MSEC 2026 (June 22–26, Penn State, USA)! This symposium will spotlight how robotics, AI, digital twins, and automation are transforming manufacturing. From cobots to mobile and reconfigurable platforms, robotics now extends beyond traditional automation to enable flexible task execution, real-time decision-making, and intelligent human-machine interaction. Our goal is to highlight cutting-edge innovations and applications, while offering a forum for researchers and practitioners to exchange ideas on adaptive, resilient, and smart manufacturing systems. 📄 Details in the attached Call for Papers. ✨ Join us in shaping the future of smart manufacturing systems!

62

1 Comment