Skip to main content

Research Partnerships & Open Challenges

Duration: 35 min · Level: Foundational · Module: 10. G1 Development Roadmap · Focus: research, partnerships, challenges, strategy

No humanoid program succeeds in isolation. The field moves fastest through open collaboration — shared benchmarks, published methods, open-source policies, and researchers who carry ideas between labs and companies. For the G1, the most relevant labs are simultaneously a knowledge base to build on and a map of who to partner with. This closing lesson surveys those groups, names the problems still unsolved as of 2030, and explains the strategic asset that turns deployment itself into a competitive advantage: the data flywheel.

The labs that matter for G1​

Four research groups do work that maps almost directly onto the G1's subsystems, and knowing which is which tells you where to look for talent, methods, and collaboration.

CMU Robotics Institute is the most directly applicable to the G1's manipulation stack — whole-body control, dexterous manipulation, and the ALOHA bimanual systems. When the question is "how should two arms and a body coordinate to do useful work," CMU's output is the first place to look.

MIT CSAIL and the Biomimetic Robotics Lab are the source of the QDD actuation concept that underpins the G1's locomotion joints (the through-line back to Lesson 2.1's Mini Cheetah lineage), and they bring locomotion, actuator design, and contact-rich manipulation — plus a strong history of collaborating with startups, which lowers the barrier to engagement.

ETH Zurich's Robotic Systems Lab (RSL) wrote most of the foundational papers the G1's locomotion stack will rely on, through their work on locomotion RL, the ANYmal platform, and sim-to-real transfer. If the G1's RL-plus-MPC approach has intellectual parents, many of them are at RSL.

UC Berkeley's RAIL anchors the open-source VLA ecosystem — Diffusion Policy, ALOHA, and RT-2 collaborations — and is accessible both through open-source code and through Berkeley researchers who routinely join companies. For a program that plans to fine-tune an existing VLA rather than train one from scratch, Berkeley's open ecosystem is the supply line.

The strategic read: the G1's architecture is deliberately built on the published, open work of these groups, which means the team's job is integration and specialization, not reinvention. Partnership and recruiting follow the same map.

The problems still open in 2030​

Honesty about what remains unsolved is what keeps a roadmap from overpromising. Three challenges are expected to persist toward 2030.

Long-horizon task planning — tasks longer than roughly 50 steps — remains hard. Current policies excel at short skills but struggle to chain dozens of them into a reliable plan without compounding errors. Any G1 task that decomposes into a long sequence inherits this fragility.

Robust outdoor locomotion in rain, wind, and genuinely uneven terrain is unsolved. Indoor locomotion is largely tractable; the open air, with its weather and unstructured ground, is a different and harder regime — which is part of why the G1's milestones are framed around indoor and hospital settings.

Full 8-hour battery life with active manipulation is still open. The endurance target from Lesson 10.1 is achievable at moderate activity but contested under continuous, demanding manipulation — the same battery-versus-performance tension flagged as a risk in Lesson 10.2, here recast as a field-wide open problem rather than a G1-specific one.

Naming these as open problems sets expectations correctly: the G1 is being designed around them, not pretending they are solved.

The data flywheel as strategy​

The G1's most durable competitive advantage is not any single subsystem — it is the compounding loop that deployment creates. Every deployed robot generates real-world training data. Put numbers to it: 100 deployed G1s, each generating 10 hours a day, produce 1000 hours of data per day. That volume of real, in-distribution experience is exactly what the sim-to-real mitigation in Lesson 10.2 and the VLA V2 milestone in Lesson 10.3 depend on — and it dramatically accelerates policy improvement.

The flywheel's logic is self-reinforcing: more deployments produce more data, which produces better policies, which make the robot more useful, which justifies more deployments. A competitor without robots in the field cannot generate this data, which is why getting units deployed early — even at limited capability — matters strategically far more than its near-term revenue suggests. The beta deployment in 2027 is, in this light, less a sales milestone than the moment the flywheel starts to turn.

Putting it into practice​

Build the collaboration-and-data strategy that complements the technical roadmap.

  1. Make a table mapping each G1 subsystem (manipulation, locomotion, actuation, VLA) to its most relevant lab — CMU, ETH RSL, MIT, Berkeley RAIL — and note whether you engage it through open-source code, recruiting, or direct collaboration.
  2. For each open problem (long-horizon planning, outdoor locomotion, 8-hour active battery), write one sentence on how the G1 design avoids or contains it today.
  3. Model the data flywheel for your own deployment scale: pick a robot count and hours-per-day and compute the daily data volume, using the 100-robots-at-10-hours-equals-1000-hours-per-day reference.
  4. Identify the single bottleneck that limits how fast the flywheel turns — likely deployment count or data infrastructure — and make accelerating it an explicit program goal.
  5. Decide which open problem is most likely to become a G1 problem first, and place a research-collaboration bet against it before it blocks a milestone.

Key takeaways​

  • Four labs map onto G1's stack: CMU (manipulation), MIT CSAIL and Biomimetic (actuation/locomotion, source of QDD), ETH RSL (locomotion RL and sim-to-real), and Berkeley RAIL (open-source VLA).
  • The G1 is built on these groups' published, open work, so the team specializes and integrates rather than reinventing — and the same map guides partnership and recruiting.
  • Three problems stay open toward 2030: long-horizon (>50-step) planning, robust outdoor locomotion in weather, and full 8-hour battery under active manipulation.
  • The data flywheel is the core strategic asset: 100 deployed G1s at 10 hours/day generate 1000 hours/day, compounding into faster policy improvement.
  • Early deployment matters more for the data it produces than for near-term revenue — the 2027 beta is where the flywheel starts turning.
  • A credible program pairs its technical roadmap with a collaboration map and a data strategy, and stays honest about what the whole field has not yet solved.

← Previous: 10.3 2026–2030 Development Milestones

Part of Module 10: G1 Development Roadmap.