Fauna Sprout: A lightweight, approachable, developer-ready humanoid robot

TL;DR

Sprout is a humanoid hardware platform from Fauna Robotics designed around safety, expressivity, and developer accessibility rather than raw capability. It is a 1.07 m / 22.7 kg, 29-DoF child-sized humanoid with soft exterior panels, backdrivable motors with conservative torque limits, and compliant whole-body control — the explicit thesis is that the field needs a robot that is cheap and safe to operate around people, not another industrial bipedal demo. The release ships an integrated stack: whole-body control, manipulation with integrated grippers, VR teleoperation, mapping/navigation, and on-board compute (Jetson AGX Orin, 64 GB RAM, 1 TB SSD) for local perception/policy inference. For Luma it is a credible reference hardware target for any embodied / VLA work that needs to leave the simulator and run around researchers.

Key claims

Sprout is a 1.07 m, 22.7 kg, 29-DoF humanoid whose form-factor scale was chosen explicitly to limit kinetic energy and potential impact forces to substantially improve the safety profile around people [§Abstract / Introduction].
Mechanical safety features are layered with the scale choice: soft exterior panels, minimized pinch points, and backdrivable motors with conservative joint torque limits [§Introduction].
At the control level, whole-body behaviors run through compliant controllers that minimize contact forces during close-proximity operation, rather than position-tracking controllers with hard reference following [§Introduction].
The platform exposes a unified hardware-software stack that combines whole-body control, manipulation with integrated grippers, and VR-based teleoperation [§Abstract].
Integrated teleoperation, mapping, and navigation capabilities are shipped on-board rather than left to the developer, with the stated goal of lowering the expertise barrier for educators, designers, and creative technologists [§Introduction].
On-board compute is a Jetson AGX Orin with 64 GB RAM and a 1 TB SSD, sized for local image recognition, voice, navigation, and policy inference without cloud dependency [reported via Fauna platform spec; cited in coverage of the paper].
Operating modes are dual: VR-headset teleoperation (operator drives the full body in real time) and modular autonomy [reported platform mode description].
Battery is swappable, with 3–3.5 hours of runtime per charge [reported platform spec].

Method

The arxiv submission is a technical report on a hardware platform, not a new learned-control or algorithmic contribution. The design choices are organized around three constraints — safety, expressivity, accessibility — and the report justifies each subsystem against them:

Mechanical: child-scale anthropomorphic body (1.07 m / 22.7 kg / 29 DoF), soft compliant exterior panels, joint travel and torque limits set to bound worst-case impact forces, pinch-point minimization in linkage design.
Actuation: backdrivable motors throughout, enabling both safety (a person can physically move the limbs) and compliant control.
Control: whole-body compliant controllers replace stiff position-tracking; manipulation goes through integrated grippers. The exact controller family is the report’s main control-stack contribution; expressive behaviors include eyebrow movement and head pose alongside body motion, treated as first-class outputs rather than afterthoughts.
Teleoperation: VR-headset full-body teleop is the primary developer-facing interface, with mapping and navigation as supporting capabilities for autonomy mode.
Compute: Jetson AGX Orin on-board (64 GB RAM, 1 TB SSD) so perception, policies, and logging stay local; swappable battery for 3–3.5 h sessions.

The “developer-ready” framing is a deliberate contrast with both closed industrial humanoids (Tesla Optimus, Figure, 1X) and academic prototypes that require deep low-level expertise to operate.

Results

There are no learned-policy benchmark numbers in this paper — it is a platform / system report, not a learning paper. The “results” are qualitative demonstrations of the design envelope, illustrated in Fig. 1: standing and looking up at a person, closed-loop high-five interaction, kneeling at a coffee table to manipulate interlocking toy bricks, reaching the ground to pick up an object, being squeezed by a child’s hand, posing with a toy flower, and dancing with expressive lights and eyebrow movements [Fig. 1]. The scenarios are chosen to span the design axes — safe close-proximity contact (high-five, child squeeze), manipulation in human spaces (toy bricks, ground pickup), expressivity (dance, eyebrows, light), and approachability (flower pose).

Why it’s interesting

For a generative-AI lab moving toward embodied work, the interesting move here is the inversion of the usual humanoid trade-off: existing humanoid programs (Optimus, Figure, 1X, Unitree, GR00T-targeted platforms) optimize payload, locomotion robustness, and adult-scale form factor, then bolt on safety; Sprout starts from “this thing has to be safe to deploy in a room with researchers” and lets that constrain everything downstream — scale, compliance, torque limits, controller family. That makes it a viable target hardware for VLA / world-model research that wants to close the sim-to-real loop with people in the loop, which is exactly the regime most generative-policy work avoids today. The 29-DoF + integrated grippers + VR teleop also gives the data-collection axis Luma would actually care about: it’s the right form factor for harvesting human-quality demonstrations rather than industrial pick-and-place trajectories.

Worth noting that the author list is unusually long (45+ authors, “Fauna Robotics” as the corporate author) — this reads as a company team-wide technical report rather than a research-team paper, comparable to NVIDIA’s GR00T platform announcements or the LTX-2 team release model.