Top 10 People Shaping AI Robotics in 2025

Artificial intelligence is finally getting hands—and grippers. The once-separate worlds of machine learning and mechatronics have collided into “AI robotics,” where models reason about the physical world and robots act in it. This article spotlights The Top 10 Influential Figures in AI Robotics Today, and—crucially—translates their playbooks into practical steps you can apply. Whether you’re a product leader, early-career roboticist, startup founder, or policy analyst, you’ll learn what each figure is known for, why their approach matters now, and how to adopt their methods with off-the-shelf tools, modest budgets, and safe practices.

Key takeaways

• AI robotics is moving from demos to deployment. Leaders are pairing large models with robust control, data engines, and safety benchmarks to make robots useful beyond the lab.
• Foundation models for robots are here. “Robotics foundation models” trained on multimodal real-world data are powering adaptable manipulation and mobile autonomy.
• Vision-language-action is an inflection. Techniques like RT-2 and Gemini Robotics show how web knowledge and embodied reasoning transfer into robot control.
• Hardware still matters. Breakthroughs in actuation, soft robotics, and safety envelopes remain essential for reliability, not just model size.
• You can get hands-on cheaply. Combine affordable kits (TurtleBot, low-cost arms), open simulators (MuJoCo, Isaac, PyBullet), and ROS 2 to reproduce many ideas at home or in a small lab.

1) Demis Hassabis — Unifying models and machines at Google DeepMind

What it is and core purpose. As CEO of Google DeepMind, Demis Hassabis steers research that merges perception, language, and action—recently showcased in Gemini Robotics (and Gemini Robotics-ER) to interpret instructions and complete multi-step physical tasks, alongside earlier VLA efforts like RT-2 that translate web knowledge into robot actions. The strategic bet: foundation models + embodied reasoning will make robots broadly useful.

Requirements / prerequisites (and low-cost alternatives).

• Skills: Python, ROS 2, PyTorch/JAX basics, and comfort with dataset curation.
• Hardware: A mobile base (TurtleBot-class) or low-cost 4–6-DoF arm; optional RGB-D camera. On a budget, use only simulation.
• Software: ROS 2 (Humble/Jazzy), MuJoCo or Isaac Sim, and a VLM or small VLA baseline.

Beginner steps (Hassabis-style VLA experiment).

1. Reproduce a VLA toy task in sim: grasp colored blocks on instruction. Use MuJoCo + ROS 2 to spawn objects and a simple arm.
2. Ground a VLM to actions: discretize actions as tokens (per RT-2 recipe) and fine-tune a small model with ~5–10k synthetic trajectories.
3. Evaluate generalization: issue novel commands (“pick the smallest red block and place it near the cup”) and log success.

Beginner modifications & progressions.

• Easier: rule-based planner with a VLM only for perception.
• Harder: add chain-of-thought action reasoning, long-horizon plans, and on-robot real-world trials.

Recommended cadence & metrics. Weekly iterations with a fixed battery of 50 evaluation prompts; track success rate, steps to completion, and collisions.

Safety & common mistakes.

• Don’t deploy unguarded LLM actions to hardware. Gate outputs with motion-planning constraints and force limits.
• Avoid overfitting to templated instructions; vary language.

Mini-plan (2–3 steps).

• Week 1: Simulate pick-place; parse natural language to symbolic goals.
• Week 2: Swap the parser with a small VLA that directly outputs actions. Validate with 50 prompts.

2) Marc Raibert — Turning agility into a research agenda

What it is and core purpose. Founder of Boston Dynamics and head of the Boston Dynamics AI Institute, Raibert’s north star is dynamic capability: balance, manipulation, and athletic behaviors that look more like animal motion than scripted robotics. The institute’s mission is to fuse machine learning with world-class hardware for reliable, useful robots. Open Robotics

Requirements / prerequisites.

• Skills: control theory (MPC, whole-body control), state estimation, reinforcement learning.
• Hardware: not everyone has a legged platform; use simulation (Isaac/MuJoCo) and low-cost manipulators to practice “dynamic manipulation.”

Beginner steps (Raibert-style dynamic challenge).

1. Build a dynamic push recovery controller for a simulated biped/arm using MuJoCo; incorporate a disturbance observer.
2. Add a learned residual policy to improve robustness on uneven terrain or to catch a sliding object.

Beginner modifications & progressions.

• Sim-only to start; later port to a real arm with a spring-loaded end effector.
• Add vision for terrain/obstacle anticipation.

Recommended cadence & metrics. 100 trials per condition; measure recovery rate, time-to-stabilize, and peak joint torques.

Safety & common mistakes.

• Over-trusting sim: add domain randomization early.
• Ignoring thermal limits in extended dynamic trials.

Mini-plan.

• Day 1–2: Implement baseline controller.
• Day 3–5: Train residual policy with disturbances; log robustness curves.

3) Daniela Rus — Soft robotics and “AI that serves”

What it is and core purpose. As MIT CSAIL’s director, Rus advances soft materials, self-knowledge in robots, and generative tools that make robot design more accessible. Her lab’s work on origami-inspired mechanisms and integrated sensing/actuation broadens where robots can go, from surgical contexts to reefs. Her leadership keeps a spotlight on useful, safe, human-aware autonomy. Wall Street Journal

Requirements / prerequisites.

• Skills: ROS 2, perception, and basic fabrication (laser-cut/3D print).
• Hardware: soft gripper kit or silicone molds, microcontrollers, and a desktop arm; low-cost alternative is a compliant 3D-printed gripper on a hobby arm.

Beginner steps (Rus-style soft gripper).

1. Fabricate a simple pneumatic gripper (3D-printed molds + silicone); instrument with a pressure sensor.
2. Control via ROS 2: open/close with closed-loop pressure; evaluate on fragile objects (eggs, chips).
3. Learn grasp success models from images + pressure traces.

Beginner modifications & progressions.

• Swap silicone durometer or finger geometry.
• Add a “self-model” network to estimate shape/pose under load.

Recommended cadence & metrics. Grasp success rate over 100 trials, mean squeezing force at failure, and time-to-grasp.

Safety & common mistakes.

• Seal leaks and over-pressurization; use a regulator with relief valve.
• Don’t ignore camera-lighting consistency when training.

Mini-plan.

• Weekend: Build the gripper and baseline controller.
• Next week: Collect 500 grasps; train a small classification model.

4) Gill Pratt — Scaling embodied AI for industry

What it is and core purpose. As CEO of the Toyota Research Institute, Gill Pratt champions embodied AI that learns many “skills” safely in the lab before tackling real homes and factories. TRI’s roadmap emphasizes data engines, skill libraries, and simulation-to-real transfer to move from dozens of skills to hundreds and beyond.

Requirements / prerequisites.

• Skills: dataset ops, teleoperation logging, and safe imitation learning.
• Hardware: a single mobile manipulator is enough; for small labs, pair a TurtleBot-class base with a 4-DoF arm and a wrist camera.

Beginner steps (Pratt-style skill library).

1. Define 10 atomic skills (wipe, pick-place, open drawer) and collect teleop demos with consistent camera views.
2. Train imitation policies per skill; add guardrails (force limits, forbidden zones).
3. Compose skills into short household routines; track success.

Beginner modifications & progressions.

• Start in Isaac or MuJoCo; then replay on hardware.
• Add language “selectors” to switch skills by instruction.

Recommended cadence & metrics. Add 2–3 new skills per week; report per-skill success and composition reliability.

Safety & common mistakes.

• Don’t skip checkout lists before live runs (estop, workspace clear).
• Avoid drifting camera intrinsics across sessions.

Mini-plan.

• Week 1: Record 100 demos → train wipe + pick-place.
• Week 2: Evaluate 200 trials; fix failure modes; add drawer open.

5) Fei-Fei Li — Vision at the center of robot understanding

What it is and core purpose. Fei-Fei Li’s influence on AI robotics stems from pioneering large-scale visual datasets like ImageNet and leading human-centered AI that links perception to action and policy. Today, her work at Stanford HAI continues to shape how robots see and how AI research serves people and the public interest. Stanford ProfilesWikipediaStanford HAI

Requirements / prerequisites.

• Skills: dataset design, labeling quality control, and data ethics.
• Hardware: none required to start; you can bootstrap with public image/video datasets before grounding to a real robot.

Beginner steps (Li-style perception upgrade).

1. Assemble a balanced dataset of your workspace objects (500–2,000 images).
2. Train a lightweight detector/segmenter; validate on edge cases (glare, occlusion).
3. Plug into ROS 2 to condition your grasp or navigation pipeline on the improved perception.

Beginner modifications & progressions.

• Start with synthetic data augmentation, then fine-tune with 10% real data.
• Add language grounding so the robot can follow verbal object references.

Recommended cadence & metrics. Weekly mAP/IoU improvements and downstream task success deltas.

Safety & common mistakes.

• Avoid dataset bias that fails on darker/lower-contrast objects; diversify backgrounds.
• Monitor for open-set misclassifications that confuse downstream policies.

Mini-plan.

• Day 1–2: Data collection and cleaning.
• Day 3–5: Train/validate; integrate with pick-place.

6) Pieter Abbeel — From robot learning research to foundation models in the wild

What it is and core purpose. A Berkeley professor and co-founder of Covariant, Abbeel has long pushed deep reinforcement and imitation learning for robot dexterity. Covariant’s RFM-1 is a notable commercial “robotics foundation model,” trained on massive multimodal warehouse data, aiming for open-ended manipulation and language-conditioned control. VCR Research

Requirements / prerequisites.

• Skills: RL/IL basics, data logging, and evaluation harnesses.
• Hardware: a single arm and a depth camera; on a budget, synthetic data + sim demos first.

Beginner steps (Abbeel-style data engine).

1. Teleop 1,000 grasps across diverse objects; log RGB-D + proprioception.
2. Train a behavior-cloned policy; then add in-context hints (few examples) to improve generalization.
3. Stress test on novel objects and lighting.

Beginner modifications & progressions.

• Start with robomimic/LeRobot baselines before bespoke architectures. GitHub
• Add any-to-any sequence modeling (images → actions) to handle wider tasks. covariant.ai

Recommended cadence & metrics. Weekly grasp success on a fixed 30-item set; measure recovery from slippage and grasp speed.

Safety & common mistakes.

• Don’t skip force/torque thresholds; protect humans and hardware.
• Beware distribution shift: keep collecting “hard negatives.”

Mini-plan.

• Week 1: Build dataset and BC baseline.
• Week 2: Add in-context learning; compare success gains.

7) Chelsea Finn — Fast adaptation via meta-learning for real robots

What it is and core purpose. Finn’s research centers on generalization and rapid adaptation for robot skills—how an agent learns new tasks quickly from limited examples. She contributed to the RT-2 line of vision-language-action (VLA) work and leads Stanford research on robot learning that bridges sim-to-real and few-shot control. Stanford Profiles

Requirements / prerequisites.

• Skills: few-shot learning, data splits for “tasks” not just “examples.”
• Hardware: sim first; then a small arm plus camera.

Beginner steps (Finn-style meta-learning).

1. Define 20 tasks (e.g., place object by color/size/shape) as training tasks; hold out 5 new tasks for meta-test.
2. Train a meta-learner that updates quickly with 5–10 demonstrations.
3. Evaluate on the held-out tasks; measure steps to competence.

Beginner modifications & progressions.

• Start with simple parameter-efficient fine-tuning; progress to full MAML-style adaptation.
• Layer on language: adapt to a new verbal constraint with two demos.

Recommended cadence & metrics. Track success after N=1, 5, 10 demos; plot adaptation curves.

Safety & common mistakes.

• Avoid leaking test tasks into meta-train.
• Keep demonstrations consistent; small inconsistencies hurt few-shot learning.

Mini-plan.

• Week 1: Sim tasks + meta-train.
• Week 2: Two new tasks with 5 demos each; compare to non-meta baseline.

8) Elon Musk — Industrial push for general-purpose humanoids

What it is and core purpose. Through Tesla’s Optimus program, Musk is driving a high-profile industrial attempt at a general-purpose humanoid. Official materials emphasize factory utility, cost curves, and reusing Tesla’s AI/autonomy stack. While timelines are debated, the project has accelerated humanoid interest across the supply chain. TeslaArs TechnicaRobots Guide

Requirements / prerequisites.

• Skills: whole-body kinematics, safety envelopes, and task decomposition.
• Hardware: you likely don’t have a humanoid. Instead, work on humanoid-adjacent stacks—perception, planning, and safety layers that transfer to bipedal form factors via sim.

Beginner steps (humanoid-adjacent pipeline).

1. Pose-aware manipulation: detect articulated objects (drawers, doors) and plan constrained motions.
2. Safety monitors: force limiting, velocity caps, and stop-lines; embed “no-go” regions around humans.
3. Task scripts: break house/factory chores into segments that a humanoid would share with your arm/base.

Beginner modifications & progressions.

• Simulate a humanoid (URDF) and validate your safety supervisor against falls or self-collision.
• Add language prompts to select tasks safely.

Recommended cadence & metrics. Mean time between safety stops; completion success on chore suites; fall/near-miss counters (sim).

Safety & common mistakes.

• Never rely on LLMs for safety gating. Always hard-enforce limits.
• Plan around contact uncertainties; assume bad friction.

Mini-plan.

• Week 1: Build a task/safety supervisor around your arm.
• Week 2: Simulate biped reach tasks; log safety interventions.

9) Rodney Brooks — Human-centered, practical autonomy

What it is and core purpose. A co-founder of iRobot and Rethink Robotics, Brooks now leads Robust.AI, focused on AI-powered warehouse/mobile robots that work with people. His current writing tempers hype with field experience and argues for systems that deliver ROI today through cognition + collaboration.

Requirements / prerequisites.

• Skills: workflow design, human factors, and “AI-in-the-loop” UIs.
• Hardware: AMRs or follow-me carts; on a budget, roll a TurtleBot base with a tablet UI.

Beginner steps (Brooks-style cobot workflow).

1. Shadow a process (e-commerce picking, kitting); document exceptions and handoffs.
2. Prototype a robot-assisted flow: robot carries, human scans; keep UI dead-simple.
3. Pilot with two operators; capture KPIs (throughput, errors, travel distance).

Beginner modifications & progressions.

• Start with “walk along” assist before autonomy.
• Add predictive routing and dynamic zone assignments.

Recommended cadence & metrics. Weekly throughputs, exception rates, near-miss logs, and operator satisfaction.

Safety & common mistakes.

• Over-automating fragile steps; aim for shared autonomy first.
• Under-investing in operator training and signage.

Mini-plan.

• Week 1: Map the floor, design the UI.
• Week 2: Pilot one aisle; expand once stable.

10) Ken Goldberg — Data, dexterity, and deployment

What it is and core purpose. UC Berkeley’s Ken Goldberg advances grasping, manipulation, and the data engines behind them (e.g., Dex-Net) and co-founded Ambi Robotics to turn research into parcel-sorting and packing systems. His work exemplifies the tight loop among simulation, synthetic data, and real-world reliability. goldberg.berkeley.eduIP Industry Research AlliancesTechCrunch

Requirements / prerequisites.

• Skills: dataset synthesis, analytic grasp metrics, active learning.
• Hardware: commodity arms are fine; start with suction and parallel-jaw grippers.

Beginner steps (Goldberg-style grasp pipeline).

1. Generate synthetic scenes with varied lighting and clutter; compute grasp candidates.
2. Train a grasp quality model; validate on 30 unseen household items.
3. Close the loop: add active learning to request human labels on uncertain cases.

Beginner modifications & progressions.

• Start suction-only; add fingered grasps later.
• Introduce parcel deformables (bags, pouches) once rigid objects are reliable.

Recommended cadence & metrics. Throughput (picks/min), success rate, and damage rate.

Safety & common mistakes.

• Ignoring elongated/transparent objects—collect targeted data.
• Overfitting to single-vendor bins and lighting.

Mini-plan.

• Week 1: Synthetic training + baseline model.
• Week 2: 1,000 real picks; iterate on failure clusters.

Quick-Start Checklist: Reproduce the Leaders’ Core Ideas

• Pick your lane: VLA (Hassabis/Finn), dynamic control (Raibert), soft robotics (Rus), skill libraries (Pratt), data engines (Abbeel/Goldberg), human-centered flows (Brooks), or humanoid safety (Musk).
• Set up tools: ROS 2 (Humble/Jazzy), MuJoCo or Isaac Sim, Gymnasium/Stable-Baselines3, PyBullet for quick physics checks. GitHub
• Hardware (optional at first): TurtleBot-class base, low-cost arm, RGB-D camera; or sim-only for weeks. pybullet.org
• Data habit: Log everything—RGB-D, joint states, force/torque, success labels.
• Safety envelope: E-stop, force/velocity caps, geofences; never run raw LLM actions on real robots.
• Weekly ritual: Add 1–3 skills or tasks; run 200 structured evals; write a one-page failure report.

Troubleshooting & Common Pitfalls

• Models that “look smart” but fail physically. Add physical priors and hard constraints; test on long-horizon tasks with clutter.
• Dataset drift between sessions. Fix camera intrinsics and lighting; include calibration snapshots per run.
• Sim-to-real gap. Use domain randomization and sensor noise models; rehearse grasps on deformables (bags, pouches) before deployment.
• Over-automation. Start with assistive flows; measure ROI before full autonomy. Robust AI
• Safety theater. Practice physical drills (estop, power loss). Track near misses like you track accuracy.

How to Measure Progress or Results

• Task success rate on a fixed, public checklist (50–200 prompts).
• Throughput and reliability (picks per minute; mean time between failures).
• Generalization (success under novel instructions/objects).
• Human factors (operator satisfaction, training time).
• Safety metrics (force limit trips, near misses, collisions).

A Simple 4-Week Starter Plan (Embodied AI Track)

Week 1: Foundations

• Install ROS 2 + simulator; bring up a simulated arm/base and run a scripted pick-place.
• Build a 50-image object dataset; train a tiny detector.

Week 2: From perception to action

• Replace scripted perception with your model; add simple natural-language parsing → goals.
• Collect 200 teleop demos across 3 skills (pick, place, wipe).

Week 3: Learn skills

• Train imitation policies for the 3 skills; compose them into a routine.
• Add guardrails: force/velocity caps and “no-go” zones.

Week 4: Generalize

• Introduce a small VLA or a language-conditioned skill selector; run 500 eval trials with adversarial prompts.
• Write a one-page “deployment memo” explaining risks, mitigations, and ROI targets.

FAQs

1. What’s the fastest way to try a VLA without big GPUs?
   Start in simulation, discretize actions as tokens, and fine-tune a small vision-language model on <10k synthetic trajectories. Keep prompts short, and evaluate on 50–100 language variations.
2. Do I need a humanoid robot to work on humanoid problems?
   No. Develop safety supervisors, perception for articulated objects, and task decompositions using an arm or mobile base; validate on humanoid URDFs in sim.
3. How do I avoid overfitting to my lab?
   Use domain randomization, multiple cameras, varied lighting, and object rotations. Keep a “foreign objects” bin that changes weekly.
4. What’s a good first gripper?
   Suction is forgiving and cheap; later add parallel jaws or a soft silicone gripper to handle deformables safely.
5. Which simulator: MuJoCo, Isaac, or PyBullet?
   MuJoCo is fast and precise for control research; Isaac excels in photorealism and digital-twin workflows; PyBullet is quick to script and great for prototyping. Use what your team can maintain. RedditGitHub
6. How many demos per skill are enough?
   Start with 50–100 quality demos per atomic skill; increase if failure analysis shows mode collapse or edge-case gaps. Compose skills into routines once per-skill success exceeds ~85%.
7. Are robotics foundation models production-ready?
   They’re progressing rapidly and already power commercial systems in constrained domains (e.g., warehouses). Expect reliability to depend on data curation, safety envelopes, and fallback routines. IEEE Spectrum
8. Can soft robotics lift real payloads?
   Yes—with the right designs. Soft/Origami mechanisms can achieve surprising strength-to-weight ratios; start with fragile-object handling where compliance is a feature.
9. What’s the biggest risk when deploying LLMs on robots?
   Hallucination and unsafe actions. Keep LLMs out of the low-level control loop; use them for high-level intent and planning gated by certified constraints.
10. How do I pick which leader’s approach to emulate?
    Match your constraints: small team + operations? Start with Brooks/Goldberg (data-driven, ROI). Research lab? Finn/Abbeel (meta-learning, RFM). Ambitious product vision? Hassabis/Pratt (VLA + skill libraries). Dynamic control? Raibert. Humanoid interest? Musk (but keep safety first). VinFuture PrizeClearpath Roboticscovariant.ai
11. Which ROS should I use in 2025?
    Stick with ROS 2 LTS (Humble) or the current stable (e.g., Jazzy) for better middleware and safety features. Medium
12. What KPIs convince stakeholders?
    Task success >90% on representative workloads, throughput parity with humans for narrow tasks, and clear safety records (zero reportable incidents), plus a payback period under 12–24 months in pilot settings.

References

• Announcing Google DeepMind, Google DeepMind blog, April 20, 2023 — https://deepmind.google/discover/blog/announcing-google-deepmind/
• Google Blog: Demis Hassabis on AI’s momentum, Google, August 11, 2025 — https://blog.google/technology/google-deepmind/ai-release-notes-podcast-demis-hassabis/
• Google DeepMind unveils Gemini Robotics and Gemini Robotics-ER, Financial Times, March 2025 — https://www.ft.com/content/f0b1dff8-8936-4e05-9e0f-b1bbbb40dc02
• Wired: Google’s Gemini Robotics reaches into the physical world, Wired, March 2025 — https://www.wired.com/story/googles-gemini-robotics-ai-model-that-reaches-into-the-physical-world
• RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control, arXiv, July 28, 2023 — https://arxiv.org/abs/2307.15818
• RT-2 project page (authors incl. Chelsea Finn), July 2023 — https://robotics-transformer2.github.io/
• Boston Dynamics AI Institute: Mission & progress, 2024 — https://www.bostondynamics.ai/news
• About the Boston Dynamics AI Institute, 2024 — https://www.bostondynamics.ai/about
• Hyundai Motor Group to Acquire Boston Dynamics, Hyundai/Press, December 11, 2020 — https://www.hyundai.com/worldwide/en/newsroom/news/hyundai-motor-group-to-acquire-boston-dynamics-0000001147
• Origami-based integration of robots that sense, decide, and act, Nature Communications, March 2023 — https://www.nature.com/articles/s41467-023-37158-9
• MIT CSAIL: Robots that know themselves, June 26, 2025 — https://www.csail.mit.edu/news/robots-know-themselves-mits-vision-based-system-teaches-machines-understand-their-bodies
• MIT CSAIL Director Daniela Rus earns the 2025 IEEE Edison Medal, MIT CSAIL, December 4, 2024 — https://cap.csail.mit.edu/members/research/mit-csail-director-and-eecs-professor-daniela-rus-earns-2025-ieee-edison-medal
• TRI’s generative AI for robotics: from 60 skills to 1,000, Toyota USA Pressroom, October 19, 2023 — https://pressroom.toyota.com/tri-unveils-advancements-ai-robotics/
• Wired: Toyota Research Institute’s housework robots, Wired, August 29, 2018 — https://www.wired.com/story/toyota-research-institute-robots