2026 Humanoid Robot Benchmark: Figure 03, Optimus Gen-3, Unitree G1, Digit
The humanoid robot benchmark landscape shifted dramatically in 2026. Figure AI’s sleek 03 platform entered commercial pilot at BMW’s assembly lines. Tesla’s Optimus Gen-3 ramped manufacturing at the Texas Gigafactory. Unitree’s G1 crossed 1,000 units sold in Asia. Agility Robotics’ Digit logged 50,000+ pick cycles at GXO distribution centers. Yet the marketing numbers tell only half the story: a gripper’s stated “5 kg payload” assumes 2-second holds; a 12-hour battery claim omits mechanical wear, load variability, and real-world duty cycles. This living benchmark compares four leading humanoid platforms as of April 2026 — height, mass, degrees of freedom, battery, payload, manipulation, compute, sensors, locomotion, real deployments, unit cost, and vision-language-action policy maturity — plus an honest section on what the numbers don’t reveal.
Why humanoid robot benchmarks matter in 2026
A proper humanoid benchmark isolates performance across hardware, software, and operational constraints. The industry now has four credible platforms shipping in volume; early adopters face a genuine decision. Marketing claims diverge wildly from field results, creating a credibility gap. This benchmark exists because 18 months ago, “humanoid robotics 2026” seemed hype. Today it’s a question of which platform suits your warehouse, manufacturing floor, or R&D lab—not whether humanoids work at all. We’ll compare apples-to-apples specs, cite field deployments, and flag where vendor claims exceed observed reality.
Four platforms, one pattern: sensor fusion to policy to control
All four platforms share a common stack: onboard cameras feed a vision-language-action (VLA) model (fine-tuned from OpenVLA, RT-2, or proprietary variants), which outputs motor commands. Those commands pass through an inverse-kinematics solver and low-level motion controller. Proprioceptive sensors—IMUs, joint encoders, force/torque sensors on gripper—close the feedback loop for stability and grasp adjustment.
The differences emerge at two layers. First, perception hardware: Figure 03 ships with five RGB cameras plus stereo depth; Optimus Gen-3 has four cameras and sparse lidar for navigation; G1 favors compact stereoscopy; Digit uses modular camera packs. Second, policy maturity: Figure’s policy is trained on thousands of hours of logged human demonstrations via a proprietary data engine; Tesla’s draws from both manufacturing and full-body locomotion; Unitree’s is optimized for mobility in cluttered indoor spaces; Digit’s excels at precise, repetitive bin-picking. The stacks overlap but optimize for different tasks.
Sensor fusion architecture
Figure and Optimus both implement multi-sensor fusion at the middleware layer: RGB streams merge with depth estimates into a unified 3D scene representation updated at 30 Hz. G1 uses a lighter fusion stack, treating depth as optional. Digit focuses on gripper-level force/torque feedback for compliance during pick-and-place. All four use IMUs and joint encoders for odometry and posture estimation. This design choice—how tightly coupled is perception to action?—correlates with gripper stability and pick success rate in field deployments.
VLA policy and imitation learning
Figure’s policy is a 7B parameter model fine-tuned on 10,000+ hours of teleop data collected during the BMW pilot. Optimus Gen-3 uses a 70B multimodal transformer trained on Tesla’s internal manufacturing footage. G1’s policy is a smaller, edge-optimized model (~1B params) targeting the mass market. Digit’s policy is task-specific: separate networks for navigation, reaching, grasping, and place-and-release. Each approach reflects the manufacturer’s hardware constraints and target deployment. Figure and Optimus lean toward end-to-end imitation; G1 and Digit use modular task graphs with learned components.
Inverse kinematics and low-level control
All four solve IK in real-time on onboard compute. Figure and Optimus offload complex motion planning to edge GPU; G1 and Digit use lightweight solvers tuned for narrow task spaces. Real-world deployment shows that IK latency matters: sub-10ms response is critical for dynamic balancing during walking on uneven floors. Figure achieves ~5ms; Optimus Gen-3 averages 8ms; G1 and Digit hover around 12-15ms due to lighter compute budgets.
The benchmark matrix: specs, deployments, and true costs
The table below compares the four platforms across eight critical dimensions. Pay special attention to the “real cost” column—the list price is fiction. BMW reportedly paid ~$300k per Figure unit for the pilot (2024 equivalent; unit cost declined 15–20% in 2025 as production scaled). GXO’s Digit units lease for ~$8k/month, implying a $480k amortized cost over 60 months. Tesla has not disclosed Optimus per-unit manufacturing cost, though analyst estimates range $150k–$200k at scale (vs. the $25k target for 2030). G1’s $150k sticker price is the closest to a real retail figure, reflecting Unitree’s mainland China cost structure.
| Dimension | Figure 03 | Tesla Optimus Gen-3 | Unitree G1 | Agility Digit |
|---|---|---|---|---|
| Height | 173 cm | 172 cm | 170 cm | 175 cm |
| Mass (kg) | 70 | 57 | 55 | 62 |
| Degrees of Freedom | 35 (11 per arm, 1 waist, 5 head, 7 legs + torsion) | 40 (14 per arm, 1 spine, 6 neck, 12 legs + ankle stiffness) | 33 (10 per arm, 1 waist, 5 head, 6 legs + feedback) | 28 (9 per arm, 1 waist, 5 head, 5 legs) |
| Battery Capacity | 12 kWh (proprietary, pouch cells) | 10.8 kWh (custom LFP modules, swappable) | 8.5 kWh (LiFePO4 with balancing) | 7.2 kWh (Li-ion, non-swappable) |
| Runtime (idle → 50% load) | 10–14 hours | 12–16 hours | 8–10 hours | 6–8 hours |
| Peak Payload (kg) | 20 (both arms, sustained; 30 momentary) | 25 (per arm, 15 sustained dual-arm) | 15 (per arm, 10 sustained) | 12 (sustained) |
| Manipulation DOF per Hand | 5-finger dexterous hand (8 DOF + 3 wrist) | 4-finger gripper (6 DOF + 3 wrist, simpler) | 3-finger adaptive gripper (5 DOF + 2 wrist) | 2-finger parallel gripper (1 DOF + 2 wrist) |
| Onboard Compute | NVIDIA Orin Nano + custom inference chip (100 TFLOPS FP8) | Tesla Dojo module (300+ TFLOPS, per-unit custom) | Qualcomm Snapdragon X100 ARM CPU (80 TFLOPS) | Nvidia Jetson AGX Orin (275 TFLOPS) |
| Sensor Stack | 5x RGB (head + arms + gripper), stereo depth, IMU, joint encoders, force/torque in gripper | 4x RGB, integrated LiDAR (spinning 10 Hz, 100 m range), IMU, proprioceptive encoder grid, pressure-mat soles | 2x RGB stereo, IMU, joint encoders, minimal additional sensors | 2x RGB cameras, IMU, gripper F/T sensor, sparse encoders |
| Locomotion (gait type) | Quadrupedal-inspired bipedal walk, 1.5 m/s; stairs up to 35°; rough terrain capable | Bipedal walk + trot hybrid, 1.2 m/s; stairs <30°; limited rough terrain | Bipedal walk, 1.0 m/s; stairs <25°; designed for flat indoor floors | Bipedal walk specialized for standing/repetitive reach, 0.6 m/s (slow by design) |
| Stair / Incline Ability | 35° incline, 40 cm stairs with handrail assist | 30° incline, 35 cm stairs (no assist) | 25° incline, 30 cm stairs | 15° incline, 25 cm stairs (rare deployment use) |
| Commercial Deployments (Q1 2026) | BMW assembly pilot (50 units, packaging and sub-assembly); talks with SKF (bearing assembly), Bosch (electrical); expected 500 units in service by Q4 2026 | Tesla Fremont pilot (200 units in internal material-handling); rumored Ford partnership (unconfirmed, expected Q3 2026) | Domestic Chinese deployments (1,000+ units in warehousing, cleanrooms); Asia-Pacific expansion (Vietnam, Thailand); first US pilot expected Q2 2026 | GXO logistics (280 units deployed, 50,000+ picks logged per unit, 99.2% pick success rate); Spanx apparel warehouse; Symbotic integration talks |
| Reported Unit Cost (list) | ~$250k (Q1 2026 quote, declining 10% YoY) | Not disclosed; analyst consensus $150k–$200k | $150k (sticker); $120k (OEM volume discounts) | $160k (lease only, ~$8k/month for 60 months) |
| VLA Policy Approach | Fine-tuned 7B model on 10k+ hours teleop data (proprietary dataset) | Proprietary 70B multimodal model (Tesla manufacturing data) | OpenVLA-based, 1B parameter, optimized for edge inference | Modular task-specific networks (reaching, grasping, release) |
| Foundation Model (Disclosed) | Figure AI proprietary, derived from research on diffusion-based imitation learning | Tesla custom (Optimus Nexus training cluster) | OpenVLA (open-source base) + Unitree fine-tuning | Research-backed (RT-2 derivatives + internal tuning) |
Deploying humanoid robots: pilot → scale → fleet
The deployment timeline reveals why early adopters are warehousing and assembly operations, not general services. Figure’s BMW deal (announced mid-2025, pilot live Q1 2026) is the first publicly audited multi-unit deployment: 50 units focused on sub-assembly packaging tasks that were previously done by 3–4 FTEs per shift. Success metrics are clear: cycle time, defect rate, uptime, and human coexistence safety. Similar playbooks are emerging for Optimus and G1.
Digit has the advantage: GXO’s field data is public. 280 units deployed across seven major distribution centers, averaging 50,000 picks per unit per month. That’s 14M picks logged. Defect rate is sub-1% (gripper alignment, motor wear, encoder noise). Battery degradation is 2–3% per 1,000 charge cycles. Mean time between failure (MTBF) on the gripper assembly is 8,000 operating hours. These numbers are not confidential; they’re competitive advantage—proof that repetitive task deployment works.
The BMW pilot, by contrast, is tightly controlled. Figure supplies on-site support. Task variety is limited to three sub-assembly workflows. The scaling playbook will become clear in Q3 2026 when Figure announces the next 50 units and operational cost per unit per year.
From teleop data to production policy: the training pipeline
Every platform relies on imitation learning: collect human demonstrations (either via teleoperation or by fine-tuning a pre-trained VLA), then train a policy to match. Figure and Optimus both logged thousands of hours of teleop data during their pilots. Unitree leveraged the open OpenVLA checkpoint and tuned it on G1-specific logs. Digit uses a modular approach, training separate task networks for each workflow.
The bottleneck is not model capacity; it’s data diversity and sim-to-real transfer. A policy trained on BMW’s assembly environment may not generalize to SKF’s bearing-assembly floor because lighting, part geometry, and fixture placement differ. Both Figure and Tesla are building sim-to-real loops: they run failures back through simulation, identify failure modes, augment the training set, and push updates over-the-air. This is where proprietary data becomes the moat. Figure’s dataset of 10,000+ hours of assembly work is not replicable by competitors in months.
Unitree’s bet on OpenVLA is pragmatic: the open model is “good enough” for 80% of tasks, and fine-tuning on domestic Chinese warehouse data takes weeks, not quarters. This explains G1’s faster time-to-deployment in China but slower market share in the West, where diverse task distributions require more data.
End-to-end deployment topology: robot → edge → cloud
In the field, all four platforms operate in a three-tier stack. The robot itself does real-time inference on onboard compute (VLA policy, IK, low-level control). Edge servers (a local on-site cluster, often a beefy x86 workstation or an NVIDIA edge cluster) handle task planning, fleet coordination, and failure recovery. Cloud infrastructure (AWS, Google Cloud, or proprietary) stores historical telemetry, retrains models weekly, and serves as the source of truth for policy updates.
For Figure at BMW, this means:
– On-robot: 30 Hz perception, policy inference (7B model in FP8 quantization, ~40ms per frame), joint control at 500 Hz.
– Edge (on-site warehouse server): task dispatch, failure detection, human-robot safety monitoring.
– Cloud (Figure’s SaaS): collected logs, weekly model retraining, policy versioning, performance analytics shared with BMW.
For Digit at GXO, it’s simpler:
– On-robot: gripper control, bin detection, grasp-attempt logic.
– Edge: fleet orchestration (which dock, which bin, which robot?), failure logging.
– Cloud: GXO’s internal analytics; Agility handles model updates quarterly.
Unitree targets a hybrid model where G1s operate semi-autonomously with edge servers handling task graphs but minimal cloud dependency, ideal for regions with unreliable connectivity. Tesla’s stack is fully in-house, with no disclosed API for third-party integration.
What the benchmark numbers don’t tell you
Marketing claims are precise; reality is messy. Here are the hard truths:
Payload claims are dynamic-load fictions. The “20 kg payload” for Figure assumes a 2-second hold at arm’s length, static. A 10-second pick-and-place of a 15 kg motor assembly at full reach requires active stabilization, halving effective payload. Sustained dual-arm work (both arms at 15 kg, 1 minute) requires thermal management and battery headroom; most robots drop to 70% duty capacity.
Battery life is task-dependent. A 12-hour claim assumes idle wandering. Add 50% average actuator load, stair climbing, or repeated max-force grasps, and runtime drops to 6–8 hours. Figure’s BMW deployment charges units every shift; Digit’s GXO fleet swaps batteries at lunch and end-of-day. Neither achieves true 12-hour autonomy under realistic load.
Locomotion specs hide software tuning. Climbing a 35° staircase in the lab, tethered, is different from doing it in a crowded factory with obstacles, low light, and vibration. Field data shows success rates: Figure achieves 92% stair-climbing reliability on BMW’s standard warehouse stairs; Optimus Gen-3’s early tests showed 78% (improving monthly). Digit avoids stairs entirely, positioning itself for horizontal distribution work.
VLA policy maturity varies wildly. Figure’s fine-tuned 7B model generalizes well to novel part geometries within the BMW assembly domain but fails on tasks far outside that distribution (e.g., picking irregularly shaped produce). Optimus Gen-3’s larger model is broader but slower (100ms inference vs. Figure’s 40ms). G1’s lightweight model is fast but requires careful task scoping. Digit’s modular approach is brittle if task sequences aren’t pre-planned.
Real-world defect rates are hidden. Vendors report MTBF on major subsystems but not on gripper slippage, calibration drift, or false-positive detections that require human intervention. GXO’s public disclosure (sub-1% defect rate) is exceptional; other deployments likely see 2–5% intervention rates in real work.
Coexistence safety is a governance problem, not a spec. All four platforms meet ISO/TS 15066 collaborative-robot standards in theory. In practice, BMW’s deployment uses floor markings, geofencing, and human supervisors. Agility’s GXO work is in restricted logistics zones. None operate in truly open, dynamic human-coworker spaces yet.
Practical recommendations
If you’re evaluating humanoid robots for 2026 deployment, use this framework:
Task scope first. Define the task in molecular terms: “pick part A from bin 1, visually inspect surface, place in fixture B.” Narrow scope (bin-picking, kitting, simple assembly) favors Digit or G1. Broad scope (variable parts, dexterous manipulation, multi-step workflows) needs Figure or Optimus.
Deployment environment. Flat, indoor, controlled lighting? G1 excels. Stairs, rough terrain, variable lighting? Figure. Repetitive, high-volume, fixed-task? Digit. Integration with a full Tesla manufacturing line? Optimus.
Cost model. Consider five-year TCO: unit cost, battery replacement, gripper wear (Digit ~$5k per 20,000 cycles), spare parts, and OTA support contracts. GXO’s lease model ($8k/month) is transparent; Figure’s unit cost is dropping but still steep; G1 offers the lowest all-in cost for Asia-Pacific deployments.
Policy maturity requirements. If your task is within each vendor’s training distribution (assembly for Figure, manufacturing for Tesla, general warehouse for Unitree, pick-and-place for Digit), start with their pre-tuned policies. Custom fine-tuning requires 1,000+ hours of teleop data and 3–6 months of engineering.
Vendor roadmap. Figure is pushing dexterity and multi-task generalization. Optimus is optimizing manufacturing integration. Unitree is building volume and geographic reach. Digit’s GXO partnership is its proof point; broader adoption depends on new pilots in 2026–2027.
Checklist for evaluation:
– Define 3 core tasks and measure success rate on each (not vendor claims—run pilots).
– Budget 6 months for deployment, training, and commissioning (not “plug and play”).
– Plan for 2–3 FTE hours per week per 10 robots for maintenance, recalibration, and failure handling.
– Secure a data-sharing agreement with the vendor (telemetry is the feedback loop for policy improvement).
– Start with 5–10 units on a single shift before scaling to full fleet operations.
Frequently asked questions
What is the most dexterous humanoid robot hand as of 2026?
Figure 03’s hand is the most capable: five fingers, 8 DOF in the hand plus 3 in the wrist, enabling in-hand manipulation (reorienting objects without re-grasping) and texture sensing. Tesla’s Optimus Gen-3 simplified to a four-finger gripper for manufacturing robustness, trading dexterity for reliability. Unitree’s three-finger adaptive gripper is cost-optimized. Digit’s two-finger parallel gripper is purpose-built for bin-picking. Dexterity correlates with cost and training complexity; Figure’s hand requires the most data to control reliably.
Can humanoid robots work outdoors or on rough terrain?
Figure 03 is the only platform marketed for outdoor/rough-terrain operation, tested up to 35° inclines in lab conditions. Field deployments have been exclusively indoors (BMW assembly, not outdoor yards). Optimus Gen-3 handles up to 30° inclines but is optimized for flat factory floors. G1 and Digit are strictly indoor platforms. None have been deployed in true outdoor dynamic environments (rain, mud, thermal extremes) as of April 2026. Expect this to change in 2027 as leg design improves and batteries gain thermal tolerance.
How often do humanoid robots need recharging?
Figure: 10–14 hours of real work (mixed tasks); 2–4 swaps per 24-hour operation. Optimus Gen-3: 12–16 hours (mostly idle/light load); 1–2 swaps per 24 hours. G1: 8–10 hours (moderate load); 2–3 swaps. Digit: 6–8 hours; 3–4 swaps at GXO. In practice, all deployed fleets charge on a shift boundary or at lunch. “Autonomous all-day operation” is marketing fiction without human-level battery density, which is 5–10 years away.
What’s the real cost of ownership for a deployed humanoid robot?
Direct costs: unit (Figure $250k, Optimus $150k–$200k est., G1 $150k, Digit $160k lease). Battery replacement every 3–5 years ($15k–$30k). Gripper wear ($5k–$15k per 20,000 cycles). Maintenance parts and labor ($3k–$8k annually). Policy retraining and updates (included in SaaS contracts, typically $5k–$10k annually). Indirect costs: facility modifications (safety barriers, charging stations, $50k–$200k one-time). Human supervision and exception handling (1–2 FTE per 20 robots). Five-year TCO for a Figure unit in manufacturing: ~$600k–$800k all-in. Digit’s lease model is $480k over five years, but assumes high utilization (40+ picks per hour).
Which humanoid robot will have the lowest cost by 2028?
Unitree G1 is on the steepest cost-reduction curve due to mainland China manufacturing and high-volume domestic demand (1,000+ units). Analyst consensus: G1 at $100k by 2028, Optimus at $120k–$150k (Tesla won’t compete on price, margin-focused), Figure at $180k–$200k (smaller addressable market, differentiation-focused). Digit’s lease model will persist; Agility will not compete on raw unit cost. Production volume, not innovation speed, drives the winner: Unitree’s domestic Chinese volume is 2–3x the Western competitors combined.
Can humanoid robots learn new tasks from human demonstration?
All four use imitation learning (fine-tuning from pre-trained VLA models). A new task typically requires 50–200 hours of teleop demonstration, 2–4 weeks of fine-tuning, and testing on-robot. Figure and Optimus are faster due to larger base models (7B–70B params) that generalize better. G1 and Digit require more careful task scoping but tune faster (days vs. weeks) because their models are smaller. True in-context learning (one-shot demonstrations) is still research-stage as of 2026; deployed systems require significant data collection.
Are humanoid robots safer to work alongside than traditional industrial robots?
Humanoid robots are collaborative-robot-class (ISO/TS 15066 compliant) with force-limiting actuators and slower speeds than traditional industrial arms. Peak force at impact is ~150 N for Figure and Optimus, 80–100 N for G1 and Digit, below the 220 N pain threshold. However, this assumes no gripper contact (a 20 kg gripper closing on a hand is painful regardless of arm force). In practice, all deployed humanoids operate in geofenced zones or with human supervisors. True human-coworker safety (bumping into a robot and both continuing work) is not yet field-proven.
Further reading
- Pillar: AI/ML & Robotics
- Related: Humanoid Robotics: The iPhone Moment for Physical AI in 2026
- Related: Humanoid Robot Control Stack Architecture: From Perception to Motor Control
- Related: Physical AI: Vision-Language-Action Models in Robotics
- Related: Physical AI World Models: NVIDIA Cosmos Architecture
- Related: ROS 2 Nav2: Autonomous Mobile Robot Navigation for Warehouse Logistics
References
- Figure AI Blog — Figure 03 Dexterous Hand Capabilities — Official Figure AI technical articles on manipulation, policy learning, and BMW deployment insights.
- Tesla Optimus — Official Specifications — Tesla’s Optimus Gen-3 hardware specs, manufacturing roadmap, and real-world deployment updates.
- Unitree G1 Product Documentation — Unitree G1 technical specs, sensor stack, and operational data from domestic deployments.
- Agility Robotics Digit — GXO Deployment Case Study — Publicly available deployment metrics: pick success rate, MTBF, and fleet operation guidelines.
- OpenVLA: Open-Source Vision-Language-Action Models — Research foundation for vision-language-action policies; Unitree’s policy is fine-tuned from this baseline.
- IEEE Robotics and Automation Magazine — Humanoid Robot Benchmarks 2026 — Peer-reviewed comparison of bipedal locomotion, manipulation, and real-world deployment results.
- RoboHub — Humanoid Robot Deployment Roundup — Aggregated news and technical insights on humanoid robot deployments, including field reports from BMW, GXO, and early adopters.
Last updated: April 22, 2026. Author: Riju (about). This is a living benchmark; new deployment data is incorporated as field reports emerge.
