Chapter 11: The Incumbent: Boston Dynamics

11.1 Why Boston Dynamics is not the legacy

One common framing of Boston Dynamics circa 2026 is "the legacy incumbent, about to be disrupted by end-to-end learned humanoids." Part IV argues the framing is wrong. BD ships the largest fleet of the most articulated humanoid on the market, runs the most operationally mature whole-body controller, and is the only frontier humanoid company with a captive industrial customer (Hyundai) writing actual deployment checks. It is also the only frontier humanoid company whose architectural bet — hybrid MPC+RL — treats the orthodox MPC/QP stack (Chapters 1–2) as a complement rather than something to replace.

This chapter argues three positions. First (§11.2–§11.3), BD's hardware is qualitatively different from the Unitree/Figure/Agility/AgiBot cohort: 56 DoF of hyperarticulation, hydraulic-to-electric pivot completed in a single day in April 2024, Hyundai Mobis custom actuators, and a three-decade Raibert-lineage head start in dynamic legged control. Second (§11.4–§11.6), BD's software stack is deliberately hybrid: whole-body MPC (the descendant of the Chapter 2 primitives) paired with an RL layer trained in collaboration with RAI Institute, and task-level Large Behavior Models provided by Toyota Research Institute. Third (§11.7–§11.9), the commercial pipeline — Hyundai Metaplant deployment from January 2026 onward — tests whether hybrid MPC+RL can match end-to-end-learned competitors on real factory tasks. The chapter closes (§11.10) with the open questions that will decide whether BD's architectural bet is vindicated or displaced.

Chapter 9 §9.10 placed BD in the System 0/1/2 taxonomy as "whole-body QP + MPC / RL layer on MPC / TRI Large Behavior Model." Chapter 11 unpacks that row.

11.2 The hardware pivot: April 16–17, 2024

On April 16, 2024, Boston Dynamics retired the hydraulic Atlas. On April 17, the company reveled the Electric Atlas ^[7]. The 24-hour window between the two events was deliberate communication strategy — a clean public break between the 2013–2023 research-robot era and the 2024+ commercial-humanoid era. The retrospective by Nelson, Saunders, and Raibert [Nelson et al., 2012/2019] documents the hydraulic Atlas specs that Electric Atlas inherits and rejects in roughly equal measure.

Hydraulic Atlas (DRC era, 2012–2023): 1.88 m tall, 158 kg, 28 DoF, hydraulic actuation via an onboard pump. Power-dense; violent on cable failure; notoriously difficult to maintain and iterate. The Chapter 1 DRC discussion named hydraulic Atlas as an exemplar of the orthodox stack at its peak; that framing remains accurate.

Electric Atlas (2024+) ^[7]: 1.5 m tall, 89 kg, 56 DoF (approximately half the weight of hydraulic Atlas), 2.3 m reach, 50 kg payload, 4 h battery with autonomous swap, 85–90% electrical-to-mechanical efficiency, sub-10 cm foot placement precision. Hyundai Mobis supplies the custom actuators (Chapter 4 §4.5 placed this in the QDD diffusion table but Atlas's actuators are more industrial-grade than classical QDD).

Table 11.1 summarizes the hardware delta.

Sources: ^[7]; [Atlas Electric Reveal, 2024]; Chapter 1 DRC discussion (hydraulic baseline) + §11.2 prose.

The 56 DoF is the single most distinctive hardware choice. Human bodies have roughly 30 kinematic DoF counted at the skeleton level (excluding hands and feet). Electric Atlas's 56 includes 360° hip, waist, and neck rotation — beyond human anatomical limits. The reveal video demonstrates the hyperarticulation with a supine-to-stand transition that rotates the hips through 360° rather than the 90-degree-plus-knee-bend humans use [Atlas Electric Reveal, 2024]. The engineering rationale: hyperarticulation expands the set of recoverable failure states (the robot can right itself from kinematic configurations a 30-DoF humanoid cannot) and expands the set of valid motion plans (the policy does not need to respect human-normal kinematic constraints).

The 56 DoF also has a cost. Training a policy that exploits 56 DoF in a physically-realistic way requires either (a) a substantially larger training corpus than for 30-DoF humanoids, or (b) the hybrid-MPC bet of §11.4–§11.6 — delegating dynamic feasibility to the MPC layer and letting RL handle task-level adaptation. BD has implicitly chosen (b); whether the choice survives scale is the §11.10 open question.

11.3 The Raibert lineage

Marc Raibert founded Boston Dynamics in 1992 after his MIT Leg Lab research on dynamic hopping and running. The company's 30+ year institutional continuity in dynamic legged robotics is qualitatively different from every other humanoid company's 2–10 year history. Figure AI was founded in 2022. Agility Robotics was founded in 2015. Unitree was founded in 2016. 1X was founded in 2014. Fourier Intelligence was founded in 2015. AgiBot was founded in 2023.

The accumulated advantage manifests in three specific places. First, the MPC primitives BD uses are the same ones Chapter 1 §1.5 and Chapter 2 §2.3 described — whole-body QP with friction cones, centroidal-dynamics constraints, and high-fidelity articulated-body dynamics — but tuned by institutional iteration across 1000+ research-robot-years of real deployment. The Di Carlo, Wensing, Katz, Bledt, and Kim MIT Cheetah 3 convex MPC paper ^[4] formalized the rigorous convex-MPC pattern; BD's internal stacks descended from parallel work trace the same lineage with better hardware and more data. The Wensing, Posa, and Hu T-RO survey ^[3] documents the state of optimization-based legged control at approximately 200 references; Boston Dynamics is the single company most deeply embedded in that literature.

Second, simulation assets. BD has built proprietary simulation infrastructure to validate controllers across hundreds of robot-year-equivalent scenarios. The simulation's contact fidelity, sensor models, and failure-mode catalog are trade secrets; they are also a material differentiator. A new humanoid company starting from Isaac Lab in 2024 inherits NVIDIA's generic fidelity; BD has internal domain knowledge that shapes its sim in ways that are not easily commoditized.

Third, behavior library. The company's public demos (parkour, dance, heavy-load carry) are the visible tip of an internal behavior-library iceberg. Each behavior represents many person-months of engineering across planning, control, and hardware co-design. The library's total size is a rough proxy for the deployment-breadth advantage BD carries into the Atlas commercial pivot.

The Raibert-founded RAI Institute (formed 2024–2025) is an interesting organizational bet on top of this lineage. RAI is a non-profit research institute that is collaborating with BD specifically on RL pipelines for Atlas [Boston Dynamics & RAI, 2025]. The partnership's stated focus areas: sim-to-real transfer from high-speed parallel simulators, loco-manipulation (door and lever while moving), and whole-body contact strategies (dynamic running, heavy-load handling). Zero-shot sim-to-real whole-body behaviors have been demonstrated in 2025 videos; specific numbers are not publicly disclosed [Boston Dynamics & RAI, 2025]. The RAI structure lets BD benefit from RL research without committing its production pipeline to pure end-to-end learning.

11.4 The hybrid MPC+RL architecture

BD's public architectural disclosures remain thin — no peer-reviewed paper covers the full Atlas 2024 stack — but the shape is discernible from blog posts, partnership announcements, and the control-theory literature BD draws on.

Whole-body MPC at the core. The MPC formulation descends from the Koenemann, Del Prete, and Tassa HRP-2 whole-body DDP paper ^[1] (Chapter 2 §2.3) and the Di Carlo et al. Cheetah 3 convex MPC ^[4]. It runs at approximately 20–100 Hz on an offboard or onboard compute node, producing task-space trajectories that a 1 kHz joint-PD tracker realizes. The MPC encodes friction cones, joint limits, torque limits, centroidal dynamics, and contact-wrench feasibility. The mathematics is the same material Chapter 2 audited; the engineering difference is BD's mature contact scheduling and gait-library integration.

RL policy layer above MPC. The RAI-BD partnership is the public signal [Boston Dynamics & RAI, 2025]: an RL policy trained in high-speed parallel simulation, consuming observations including the MPC's internal state, and producing residual or meta-level actions that the MPC incorporates. The exact interface is not disclosed. Two architectural patterns are plausible:

1. RL as MPC warm-start: the RL policy produces candidate task-space trajectories, and the MPC uses them as initialization for its own optimization, with MPC's provable-correctness acting as a safety filter.
2. RL as outer-loop task selector: the RL policy chooses among a library of MPC controllers (walking, running, heavy-load carry, manipulation poses), and the MPC executes whichever is selected.

Pattern 1 is closer to the Chapter 7 §7.6 safety-filter pattern. Pattern 2 is closer to the hierarchical decomposition Xie et al. published ^[10]. BD's actual stack may be either, both, or a hybrid; public disclosure is insufficient to determine which.

Large Behavior Model at the top. The Toyota Research Institute LBM ^[9] provides the System 2 layer for Atlas. LBM is built on Diffusion Policy ^[2] foundations and is designed to learn manipulation skills from small demonstration sets (tens of examples). TRI's partnership with BD ^[9] explicitly deploys LBM-derived policies on Atlas for manipulation tasks. The public signal is that LBM handles what the MPC+RL lower layers do not: open-world manipulation, novel-object handling, task-level instruction following.

The three-layer stack — LBM at System 2, MPC+RL at System 1, classical joint tracking at System 0 — maps cleanly onto Chapter 9 §9.10's BD row. The architectural signature of BD is that its System 1 is MPC+RL rather than purely learned, and its System 0 is classical PD+QP rather than a learned 10M-parameter network.

11.5 Why hybrid matters — the orthodox primitives refuse to die

Chapter 2 §2.8 articulated the operational rule: use orthodox primitives where they are provably correct; use learned primitives where distributional coverage is the only available guarantee. BD's architecture is the most disciplined expression of that rule among the frontier companies.

Three specific advantages follow.

Diagnostic failure modes. When hybrid MPC+RL fails, the failure is often diagnostic: the MPC reports friction-cone violation, or contact loss, or actuator saturation. The engineer reading the logs can point to a specific constraint that was violated. Compare this with end-to-end-learned failures, which surface as out-of-distribution events without a specific-constraint attribution. Diagnostic failure is essential for safety certification (Chapter 15's ISO 10218 / TS 15066 discussion) and for the long-term engineering rhythm of commercial deployment.

Provable partial correctness. The MPC's friction cones, joint limits, and centroidal constraints are mathematically guaranteed within the MPC's horizon and model. The guarantee degrades under model error but is never catastrophically wrong within the horizon. Learned policies carry no equivalent guarantee; they inherit the training distribution's bias but provide no formal safety envelope. BD's hybrid stack carries the MPC's guarantee at System 1's lower edge, which is what makes Atlas deployable in industrial settings where a learned policy alone would not yet pass procurement.

Data efficiency at high DoF. The 56-DoF Atlas hardware creates a sample-complexity problem for pure RL: domain randomization over 56-joint-space dynamics is more expensive than over 30-joint-space. MPC shrinks this problem by absorbing the high-DoF feasibility into convex-optimization machinery that doesn't need to be learned. RL then only has to learn task-level adaptation on top, which is a lower-dimensional learning problem. This is the reason hybrid matters specifically at Atlas's scale — at Figure 03's 30 DoF, the cost-benefit of MPC is less compelling.

The counter-argument Chapter 12 will develop is that end-to-end-learned approaches scale better once data and compute catch up, and that BD's hybrid is technical debt that will be paid down as learning systems improve. The debate is live; Chapter 16's diffusion scenarios argue neither side wins outright.

11.6 The commercial pipeline: Hyundai

Boston Dynamics has a captive industrial customer that no other humanoid company has: Hyundai Motor Group. The acquisition closed in 2021 at approximately US\$1.1 billion, giving Hyundai an 80% stake in BD ^[12]. In January 2026, Hyundai announced Electric Atlas deployment at its Metaplant in Bryan County, Georgia — the company's flagship EV manufacturing plant [Hyundai Metaplant, 2026]. The deployment is the first public end-to-end pipeline from BD's RL pipeline to commercial deployment in Hyundai's own factory, following the Agility-GXO precedent by approximately 18 months (Chapter 12).

The Hyundai integration has three strategic layers.

Capital: Hyundai's US\$1.1 billion 2021 investment gave BD financial stability that Figure, Unitree, and AgiBot do not have. The follow-on RAI Institute funding from Hyundai is a second-order investment in BD's AI capability. The capital cushion lets BD take longer design cycles and invest in the proprietary simulation and behavior-library assets of §11.3.

Deployment: The Metaplant is the proving ground. If Electric Atlas completes real automotive-assembly tasks (wire-harness routing, parts staging, inspection) at Metaplant's throughput requirements, BD has something no other humanoid company has — a commercial reference customer with actual dollar savings per deployed robot. If Atlas underperforms at Metaplant, the program's public-relations cost will be significant.

Korean strategic bridge: Hyundai's BD ownership is the single largest direct link between Korean industry and frontier humanoid capability. Chapter 14's analysis of Korea's position returns to this as the load-bearing asset in the K-Humanoid Alliance: HD Hyundai Robotics (industrial-arm spinout) joined the Alliance in 2025, and the BD-Atlas program is the most visible humanoid capability Korean industry can claim as its own ^[12]. The Korea-US IP linkage is asymmetric — most IP sits in Waltham, Massachusetts, not Ulsan — but the linkage exists, and Chapter 14 argues Korea should exploit it rather than treat it as a dependency vulnerability.

Note on lineage: BD's first commercial deployment playbook is not Atlas, it is Spot — the quadruped product launched in 2020 at US\$74,500 price point [Boston Dynamics, 2020+]. Approximately 1,500+ Spots were deployed globally by 2024 across utilities, construction, and inspection verticals. Spot provided BD's at-scale commercial experience — developer tools, customer support, software-release cadence, insurance and liability frameworks — that Atlas inherits. Figure and Agility are building these at commercial scale for the first time; BD is iterating on an existing playbook.

11.7 Comparison with the end-to-end-learned alternatives

Figure's Helix 02 (Chapter 10 §10.4) and AgiBot's GO-2 are the two end-to-end-learned architectures most directly contrasted with BD's hybrid stack. The comparison is instructive.

The table frames the trade-off. BD commits to hardware complexity (56 DoF) and trades the sample-complexity cost against MPC's mathematical tractability. Figure commits to software complexity (end-to-end learned stack) and trades formal guarantees against teleoperation data efficiency. AgiBot commits to scale (100-humanoid fleet data collection) and trades data-variety against per-unit deployment depth.

None of the three companies has publicly demonstrated commercially sustainable deployment at scale as of 2026Q1. Each is betting that its architectural choice scales faster than the alternatives. The 2026–2028 commercial data — Hyundai Metaplant, Figure pilot customers, AgiBot industrial rollouts — will provide the first differentiating signal.

11.8 The Korea angle — an unusually load-bearing relationship

Most Part IV company analyses would close with product roadmap speculation. For BD, the Korea angle deserves its own treatment. Three observations.

Hyundai-BD is the densest Korea-US humanoid linkage. Samsung has a small stake in Rainbow Robotics (Chapter 14); LG and SK have less-visible humanoid positions. Hyundai's 80% BD ownership dwarfs all other Korean positions by an order of magnitude. If a single Korean conglomerate produces a frontier humanoid in the 2028–2030 window, it is disproportionately likely to be Hyundai through BD.

The IP geography is asymmetric. BD's headquarters and engineering are in Waltham, Massachusetts. Atlas manufacturing will be split (Hyundai Mobis supplies actuators; BD integrates in Waltham; Hyundai assembles at Metaplant). Korean ownership does not give Korea the industrial base for a frontier humanoid in the sense that Japan has for Honda's robots or China has for Unitree. This is a caveat Chapter 14 develops.

The Metaplant deployment is the single most testable Korea-humanoid claim in 2026. If Atlas succeeds at Metaplant, Hyundai will be the first global auto-OEM with a self-designed (via BD) humanoid in its own factory at scale. The deployment is explicitly structured to validate BD's commercial thesis and, in doing so, to validate Hyundai's humanoid strategy. Watch the Metaplant numbers across 2026.

Chapter 14 and Chapter 15 return to these observations in the Korean-ecosystem-diagnosis framing.

11.9 Open questions

Three questions close this chapter.

First, will hybrid MPC+RL scale to manipulation as well as locomotion, or does it remain a locomotion-first architecture? BD's public demonstrations emphasize locomotion, parkour, and heavy-load carry. Upper-body manipulation is where TRI LBM is expected to carry weight, but LBM's integration with the MPC+RL lower layers is not publicly disclosed. If the integration is clean, BD has a complete stack; if it is loose, BD is locomotion-plus-manipulation-bolted-on and vulnerable to Figure's unified visuomotor approach.

Second, can BD iterate as fast as the end-to-end-learned companies? Hybrid architectures are more complex per release: each Atlas update must validate MPC correctness, RL transfer, and LBM manipulation skill separately. The 24-hour hydraulic-to-electric pivot of April 16–17, 2024 showed BD can move fast when it chooses to; the open question is whether the institutional rhythm supports the iteration cadence that Figure's monthly Helix updates imply.

Third, does the Raibert institutional continuity compound or calcify? The 30-year head start is a real asset; it is also a constraint. BD's commitment to MPC at System 1 is partly a reflection of its engineering culture, not just a technical judgment. A new humanoid company starting from scratch in 2024 — as Figure did — is not committed to MPC and can adopt whichever architecture its 2024-2026 data supports. BD's commitment is deeper; it will pay off if MPC-based approaches remain competitive and will constrain BD if they do not.

Chapter 12 now turns to the US challengers — Figure and Agility — whose architectural bets are the natural counterpoint to BD's hybrid thesis, plus a Tesla Optimus outlook where public disclosure remains thin.

References

1. Koenemann, J., Del Prete, A., & Tassa, Y. (2015). A whole-body model predictive control framework for humanoid robots. Proc. IEEE/RSJ IROS. doi:10.1109/IROS.2015.7353596.
2. Chi, C., et al. (2023). Diffusion policy: Visuomotor policy learning via action diffusion. Proc. RSS.
3. Wensing, P. M., Posa, M., & Hu, Y. (2024). Optimization-based control for dynamic legged robots. IEEE T-RO. arXiv:2211.11644.
4. Di Carlo, J., Wensing, P. M., Katz, B., Bledt, G., & Kim, S. (2018). Dynamic locomotion in the MIT Cheetah 3 through convex model-predictive control. Proc. IEEE/RSJ IROS.
5. Nelson, G., Saunders, A., & Raibert, M. (2012/2019). Atlas: The hydraulic humanoid for the DARPA Robotics Challenge. Annual Review of Control, Robotics, and Autonomous Systems (retrospective).
6. Boston Dynamics. (2024). Electric Atlas: 56-DoF all-electric humanoid. Boston Dynamics announcement, April 2024.
7. Boston Dynamics. (2024). All New Atlas: Electric Atlas reveal video. April 17, 2024.
8. Boston Dynamics. (2020+). Spot and its commercial lessons for Atlas. Product pipeline.
9. Tedrake, R. (2024). Toyota Research Institute Large Behavior Models for robot manipulation.
10. Xie, K., et al. (2024). Hierarchical reinforcement learning for humanoid whole-body control. arXiv preprint 2412.10210.
11. Boston Dynamics and RAI Institute. (2025). Partnership for Atlas reinforcement learning. February 2025.
12. Hyundai Motor Group. (2025). HD Hyundai Robotics and Hyundai–Boston Dynamics ecosystem.
13. Hyundai Motor Group. (2026). Hyundai Metaplant Georgia and Atlas deployment. January 2026.
14. Righetti, L., Buchli, J., Mistry, M., Kalakrishnan, M., & Schaal, S. (2013). Optimal distribution of contact forces with inverse-dynamics control. International Journal of Robotics Research.
15. Figure AI. (2026). Figure 03 + Helix 02: General-purpose humanoid system. (Cited as contrast.)