Category 5: Robotics -- The Physical Intelligence Frontier
Foundation models for physical manipulation are emerging just as humanoid hardware matures. The same scaling laws that powered language models are now being applied to robot control via Vision-Language-Action (VLA) models. Companies that crack general-purpose robotic intelligence will create enormous value.
Bay Area focus: 3 companies in the main section. 7 non-Bay Area companies are archived below for reference.
Bay Area Companies (3)
Click to jump to a company
Key People
Brett Adcock
Product: Helix Foundation Model + Figure 02
- Helix -- end-to-end Vision-Language-Action (VLA) foundation model built fully in-house (previously partnered with OpenAI for multimodal, now 100% internal AI stack)
- VLA architecture: takes RGB camera images + natural language instruction as input, outputs joint torques directly -- no separate perception/planning/control pipeline
- Figure 02 humanoid: 5'6" tall, 60 kg, 16 degrees-of-freedom hands for dexterous manipulation, full bipedal locomotion
- 60-minute autonomous logistics demo: Figure 02 autonomously performed warehouse pick-and-place tasks for a continuous hour without human intervention
- End-to-end neural network replaces classical robotics stack: perception, planning, and control are a single learned model
- Training pipeline: large-scale imitation learning from teleoperation demonstrations + sim-to-real transfer
- Targeting commercial deployment in BMW manufacturing and logistics warehouses
Notable Investors
Compensation (ML/Robotics Engineer)
- Median total comp: ~$275K (base + equity)
- 75th percentile: ~$400K
- Equity is pre-IPO at $39B valuation -- significant upside if robotics thesis plays out
- Source: levels.fyi, Glassdoor estimates
Why It Matters
Figure is at the intersection of foundation models and physical intelligence. Their pivot from OpenAI partnership to fully in-house AI (Helix) shows they are serious about owning the core technology. The end-to-end VLA approach (RGB + language -> joint torques) is the same paradigm shift that happened in NLP and vision. The 60-minute autonomous demo is the strongest public proof of autonomous humanoid capability. Sunnyvale HQ is ideal for Bay Area.
Path to Entry
- Roles in ML/AI for robot perception, manipulation, and VLA model training
- Strong need for engineers who understand both ML infrastructure and real-time systems
- Meta ML experience is directly relevant -- scaling, distributed training, model optimization
- Sunnyvale location aligns well
- Fast-growing -- many open positions as they scale toward commercial deployment
Tracking
Physical Intelligence Bay Area
Key People
Karol Hausman
Sergey Levine
Chelsea Finn
Brian Ichter
Product: pi0 Foundation Model
- pi0 -- transformer-based Vision-Language-Action (VLA) model, 3-5 billion parameters
- Pre-training: trained on internet-scale vision-language data (like a VLM), then fine-tuned on diverse robot demonstration datasets across multiple robot embodiments
- Input: RGB-D camera images + natural language task instruction
- Output: robot-specific joint commands (works across different robot hardware -- arms, hands, mobile manipulators)
- RECAP training pipeline: (1) Learn from human demonstrations, (2) Coaching corrections -- human provides corrective feedback on robot attempts, (3) Autonomous practice -- robot self-improves via trial-and-error with learned reward signal
- Hardware-agnostic: the same model weights transfer across different robot form factors -- single brain, many bodies
- Open-sourced pi0 in February 2025 -- released model weights and training code to the research community
- Founded by the team that created RT-1, RT-2, SayCan at Google -- the defining work in robotic foundation models
Compensation (Research/ML Engineer)
- Estimated total comp: $400K-$800K (no public levels.fyi data available)
- Elite ~50-person team of world-class researchers -- comp must compete with top AI labs (Anthropic, OpenAI, DeepMind)
- $400M raised with ~50 employees = ~$8M resources per person -- suggests premium compensation
- Pre-IPO equity at $5.6B valuation with significant upside
Why It Matters
Physical Intelligence may be the most intellectually exciting robotics company. Their founding team literally wrote the playbook on foundation models for robotics at Google. The pi0 model is the robotic equivalent of GPT -- a general-purpose VLA that can control any robot. The RECAP training pipeline (demonstrations -> coaching -> autonomous practice) is an elegant approach to scaling robot learning. $400M raised with a ~50-person team means exceptional resources per researcher. Open-sourcing pi0 shows confidence and could establish them as the standard. SF location is ideal.
Path to Entry
- Extremely selective -- hiring world-class researchers and engineers only
- Strong publications in robot learning, RL, or foundation models would be ideal
- ML infrastructure experience from Meta is very relevant -- they need to scale training for billion-parameter VLA models
- San Francisco HQ
- Network through Stanford/Berkeley robotics community
- Potentially the best learning environment for understanding how to build general AI systems
Tracking
Covariant / Amazon Robotics Bay Area
Key People (now at Amazon)
Pieter Abbeel
Peter Chen
Rocky Duan
Product: RFM-1 (Robotic Foundation Model)
- RFM-1 -- Robotic Foundation Model for manipulation, trained on billions of real robotic interactions collected from Covariant's deployed fleet
- Self-supervised learning on robotic data: unlike VLA models pre-trained on internet images, RFM-1 was trained primarily on actual robot grasping/manipulation data at massive scale
- Novel object handling: can pick and manipulate objects it has never seen before -- generalizes to new SKUs without retraining using learned geometric and physical priors
- Amazon deployment: Amazon now deploying generative AI powered by Covariant technology across its fleet of 750K+ warehouse robots
- Handles the "long tail" problem in warehouse fulfillment -- millions of unique products with varying shapes, materials, and packaging
- Commercially proven at scale before acquisition -- deployed in warehouses of major retailers for automated order fulfillment
- One of the first companies to demonstrate that foundation model scaling works for robotics in production
Compensation (Amazon Robotics -- ML Engineer)
- L5 (SDE II equivalent): ~$350K-$400K total comp (base ~$180K + RSU + sign-on)
- L6 (Senior SDE equivalent): ~$500K-$600K total comp
- Amazon RSUs vest over 4 years with back-loaded schedule (5/15/40/40)
- Source: levels.fyi Amazon data, confirmed by Blind reports
Why It Matters
Covariant proved that robotic foundation models work commercially, then got acquired by Amazon. Pieter Abbeel at Amazon Robotics means Amazon now has the strongest robotic AI team within Big Tech. With 750K+ robots and the world's largest logistics operation, Amazon Robotics offers unmatched scale for deploying robotic AI. The Berkeley office keeps this Bay Area accessible. For someone wanting to work on foundation models for robotics at massive real-world scale, this is the strongest option within a large organization.
Path to Entry
- Path is through Amazon Robotics -- standard Amazon hiring process (Leadership Principles)
- Covariant team based in Berkeley/Bay Area
- Pieter Abbeel's presence makes this the strongest robotics AI team within Big Tech
- L5/L6 ML Engineer or Applied Scientist roles
- Amazon interview process is well-documented -- prepare for LP + system design + ML depth
- Unique combination: Big Tech stability + startup-caliber research team + massive deployment scale
Tracking
Archived -- Not Bay Area (7 companies)
▶Tesla Optimus Austin, TX
Why Archived
Tesla HQ is in Austin, TX. While Tesla has a Palo Alto office, the core Optimus/AI team is primarily Austin-based. Optimus Gen 2 humanoid leverages Tesla's FSD neural network stack (end-to-end vision transformers, Dojo training cluster). Massive fleet data advantage from billions of miles of driving video. However, Bay Area presence for Optimus ML work is limited.
Boston Dynamics Waltham, MA
Why Archived
HQ in Waltham, Massachusetts. East Coast only. Products: Spot (deployed quadruped), new electric Atlas humanoid, Stretch (warehouse). 30+ years of robotics engineering expertise, now integrating AI/ML for perception and control. Hyundai backing provides manufacturing scale. Strong company but requires relocation to East Coast.
1X Technologies Moss, Norway
Why Archived
HQ in Moss, Norway. NEO bipedal humanoid designed for safe home interaction with compliant actuators. OpenAI partnership provides frontier AI access. End-to-end learned control policies. Interesting technically but requires international relocation.
Skild AI Pittsburgh, PA
Why Archived
HQ in Pittsburgh (CMU origin). Has an SF office but primary operations and founding team are in Pittsburgh. "Skild Brain" -- single foundation model controlling any robot form factor. $30M revenue in months, deployed across security robots, delivery bots, and warehouse automation. CMU founders Deepak Pathak and Abhinav Gupta. NOTE: Has SF office -- could potentially revisit if ML team is SF-based.
Sanctuary AI Vancouver, Canada
Why Archived
HQ in Vancouver, Canada. Carbon humanoid with 20+ DOF hands. Founded by Geordie Rose (also founded D-Wave quantum computing). Focus on dexterous manipulation and general-purpose cognition. Piloting with retail partners. Requires international relocation.
Agility Robotics Corvallis, OR
Why Archived
HQ in Corvallis, Oregon. Most commercially deployed humanoid robot company. ~100 Digit units in Amazon fulfillment centers. Built RoboFab (first humanoid robot factory). Oregon State University spinout. RL for dynamic walking + sim-to-real. Strong execution but requires Oregon relocation.
Apptronik Austin, TX
Why Archived
HQ in Austin, Texas. Apollo humanoid for logistics and manufacturing. UT Austin spinout with $935M+ raised. Backed by Google, Mercedes-Benz, John Deere. Modular force-controlled actuators for safe human interaction. Strong investor roster but Austin-based.