Hardware Taxonomy

The machines you can point a policy at in 2026 split into five families: arms, humanoids, hands, tactile sensors, and the compute that closes the control loop. The price spread is the story. A self-build SO-101 arm costs about $122 in parts TheRobotStudio 2026↓SO-ARM100: Low-Cost Robot Arms for EveryoneTheRobotStudio, 2026. A Unitree H2 humanoid ships for $29,900 Unitree Robotics 2026↓Unitree H2 Product PageUnitree Robotics, 2026. Boston Dynamics will not sell you an Atlas at any price, and Tesla has no Optimus 3 price to quote because there is no product to buy RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026. Between those poles sits an actual market with dated prices, disclosed degrees of freedomdegrees of freedomThe number of independent coordinates needed to specify a mechanism's configuration. An arm's degree-of-freedom count is its number of independent joints, so a 7-DoF arm places its end effector with one coordinate to spare beyond the six a rigid pose needs, and that redundancy is what lets the elbow reconfigure while the hand stays put. More degrees of freedom buy dexterity and obstacle avoidance at the price of a larger control problem., and sourcing states you can plan around. This module collects the market into one filterable guide of 40 entries, and every number that nobody publishes stays visibly empty rather than estimated.

The buyer's guide

Three reading rules. Filters compose, so "preorder humanoids in the $1k-$10k band" is two clicks; the table answers it with exactly one row, the Bumi presale. Switch the category filter to Hands and the readout falls to five entries with one price between them, which is the standalone hand market rather than a gap in the table. Price buckets use the low end of supported USD ranges. ALOHA 2 instead sits under “not disclosed”: the community estimate does not establish a configuration-specific USD list or quoted price. And every listed price carries an as-of date, because hardware pricing moved several times in 2026 alone, and a price without a date is already stale. A "not disclosed" cell marks a figure the source does not publish, and those cells sort last and filter separately by design.

Arms, from $122 to $70,000

The low end belongs to the LeRobot ecosystem. The SO-101 is a 3D-printed arm with six degrees of freedom, five joints plus a gripper, built from STS3215 servos; the community BOM for one follower arm totals $121.94 in the US column, $110.94 of that in core parts (six STS3215 servos, control board, cable, and power supply) before the clamp and tools TheRobotStudio 2026↓SO-ARM100: Low-Cost Robot Arms for EveryoneTheRobotStudio, 2026. Seeed Studio sells the SO-ARM101 Pro line around it: $295 unassembled, $299 assembled, both with 12-bit magnetic encoders and a 500 g payload Studio 2026↓SO-ARM101 Pro KitsSeeed Studio, 2026. The older Koch v1.1 design still circulates at $250-$300 on Dynamixel servos alpibrusl 2026↓LeRobot ecosystem and hardware pricing (June 2026)alpibrusl, 2026. The LeRobot software stack runs ACT, π0, and SmolVLA policies on all of them, which is what makes a $299 kit a research instrument instead of a toy Face 2026↓LeRobot DocumentationHugging Face, 2026.

A tier up, Trossen Robotics rebranded its ALOHA line as Trossen AI in 2025-2026. The WidowX AI single-arm teleopteleoperationDirect remote control of a robot by a human operator, used in robot learning mostly to record expert demonstrations. The ALOHA rig made the setup cheap and precise: the operator back-drives a leader arm that is a joint-for-joint twin of the follower the policy will run on, so the recorded motion is already in the robot's own action space. Mobile ALOHA extended the same leader-follower scheme to a mobile base for whole-body tasks. rig lists at $4,545.95 with a 1.5 kg payload and 700 mm reach across six degrees of freedom, and the Mobile AI bimanual-on-wheels rig at $33,695.95, with the Solo AI and Stationary AI at $11,385.95 and $23,995.95 between them Trossen Robotics 2026↓Trossen AI Product Line (formerly ALOHA)Trossen Robotics, 2026. All of them run 500 Hz CAN FD control with the iNerve board and integrate LeRobot and OpenPI Trossen Robotics 2026↓Trossen AI Product Line (formerly ALOHA)Trossen Robotics, 2026. The ALOHA 2 itself, the bimanual rig behind ACT Zhao 2023↓Learning Fine-Grained Bimanual Manipulation with Low-Cost HardwareTony Z. Zhao, Vikash Kumar, Sergey Levine et al., RSS 2023 and Mobile ALOHA Fu 2024↓Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body TeleoperationZipeng Fu, Tony Z. Zhao, Chelsea Finn, 2024, has no configuration-specific USD quote established by the cited source. The June 2026 community issue groups “ALOHA / ALOHA 2” at about $17k-32k without itemizing currency code, configurations, or included hardware alpibrusl 2026↓LeRobot ecosystem and hardware pricing (June 2026)alpibrusl, 2026. At the top of the category, Pollen Robotics' Reachy 2 humanoid torso goes for about $70,000 alpibrusl 2026↓LeRobot ecosystem and hardware pricing (June 2026)alpibrusl, 2026.

One arm in the guide carries no price at all: the Franka Emika Panda, the seven-DoF platform behind DROID's 76,000 trajectories Khazatsky 2024↓DROID: A Large-Scale In-The-Wild Robot Manipulation DatasetAlexander Khazatsky, Karl Pertsch, Suraj Nair, 2024. It remains the research workhorse, and sourcing it is a conversation rather than a checkout.

Photo: Ims / Wikimedia Commons. Licence: CC BY-SA 4.0.

Humanoids: shipping beat shouting

Omdia counts about 13,000 humanoids shipped in 2025 RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026. The dispute over who led says everything about the market's maturity: Omdia puts AgiBot first at 5,168 units and a 39% share, while Unitree self-reports more than 5,500 and disputes the ranking RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026. Read both numbers as an order of magnitude.

Unitree owns the priced ladder. The R1 education robot runs $4,900-$5,900 across configurations. The G1 lists at $13,500 for the base model with 23 DoF; the EDU configuration spans 23-43 DoF and is available by quote Unitree Robotics 2026↓Unitree G1 Product PageUnitree Robotics, 2026. It is also one of the two validation platforms for NVIDIA's GR00T line, alongside AgiBot's Genie-1 RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026. The H2 ships at $29,900 with 31 DoF (six per leg, seven per arm, three waist, two head), 360 N·m leg joints, and a 2,070 TOPS onboard chip Unitree Robotics 2026↓Unitree H2 Product PageUnitree Robotics, 2026. 1X prices the NEO at $20,000 outright or $499 a month against a $200 refundable deposit, with US deliveries promised by the end of 2026; the order page lists 22 DoF per hand, and a 25-DoF revision was announced on July 9, 2026 1X Technologies 2026↓NEO Home Robot1X Technologies, 2026 1X 2026↓1X Launches World Model Lab to Scale Humanoid Intelligence1X, 2026. Noetix's Bumi opened presale at ¥9,998, about $1,400 at July 2026 conversion, with delivery still unconfirmed; EngineAI's SE01 launched at ¥88K (about $12,000), Fourier's GR-3 prices above ¥200K (about $27,500), and LimX's Oli takes orders from $21,800 RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026.

The flagship tier publishes specs but not prices. Boston Dynamics' electric Atlas carries 56 DoF with continuous rotation at key joints, a 50 kg instant payload, IP67, and a 4-hour swappable battery, and every 2026 build is already committed to Hyundai and Google DeepMind Dynamics 2026↓Atlas Humanoid RobotBoston Dynamics, 2026. Figure 03 is in fleet production with an embedded palm camera per hand and 2 kW wireless charging through its feet, and Figure builds it for high-volume manufacturing at its BotQ facility Figure AI 2025↓Introducing Figure 03Figure AI, 2025. Agility sells Digit as a service with a 16 kg payload and deployments at GXO; UBTECH's Walker S2 hot-swaps its own batteries in about three minutes; AgiBot's A2 lists 40+ DoF; and Tesla's Optimus 3 has neither a price nor a current spec sheet RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026.

The shakeout around them is real. K-Scale Labs shut down in November 2025, Cartwheel Robotics in February 2026, Sanctuary AI pivoted to software in June, and Amazon acquired Fauna Robotics in March RoboZaps 2026↓38 Best Humanoid Robots in 2026RoboZaps, 2026. The companies still standing are the ones with shipping robots in this table.

Photo: DARPA / Wikimedia Commons. Licence: Public domain.

Hands ship bundled, not shelved

Dexterous hands almost never sell as standalone products; they ride along with the robot. The open exception is the LEAP Hand from CMU: 16 DoF, buildable in about four hours from catalog parts at a cost of $2,000, roughly one eighth the price of the Allegro Hand it was benchmarked against Shaw 2023↓LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot LearningKenneth Shaw, Ananye Agarwal, Deepak Pathak, RSS 2023.

Everywhere else, hand specs surface inside robot listings. The NEO's hands carry 22 DoF each, with the 25-DoF revision on the way 1X Technologies 2026↓NEO Home Robot1X Technologies, 2026. Unitree's Dex3-1, an option on the G1 EDU, has seven active DoF per hand (three thumb, two index, two middle) with an optional tactile array Unitree Robotics 2026↓Unitree G1 Product PageUnitree Robotics, 2026. The H2's dexterous hands are EDU-only options in several models Unitree Robotics 2026↓Unitree H2 Product PageUnitree Robotics, 2026. Figure 03 integrates a palm camera into each hand but discloses no hand DoF Figure AI 2025↓Introducing Figure 03Figure AI, 2025. If you want a hand to mount on your own arm in 2026, LEAP is essentially the whole catalog.

Sensors: touch is the missing channel

The datasets that generalist policies train on ship vision and proprioceptionproprioceptionSensing of the body's own state: joint positions and velocities, orientation, and contact, as opposed to exteroception, which senses the outside world through cameras or lidar. Proprioception is fast, cheap, and never occluded, and the ETH Zurich line showed how far it goes: Lee and colleagues trained an ANYmal controller that hiked mud, snow, rubble, and vegetation on proprioception alone, and Miki and colleagues kept it as the trusted channel the policy falls back on when its terrain map disagrees with its body.; DROID's rig is a Franka with three cameras and no touch sensing at all Khazatsky 2024↓DROID: A Large-Scale In-The-Wild Robot Manipulation DatasetAlexander Khazatsky, Karl Pertsch, Suraj Nair, 2024. AgiBot World is the counterexample worth naming: its humanoids carry visuo-tactile sensors alongside the camera rig, which makes it one of the few million-trajectory corpora with a touch channel AgiBot-World-Contributors 2025↓AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied SystemsAgiBot-World-Contributors, Qingwen Bu, Jisong Cai et al., 2025. The hardware is not the reason the channel stayed rare. A 2025 outlook in IEEE Transactions on Robotics puts the blame on structure: divergent transduction methods with no standardized evaluation framework, durability failures in real conditions, and temperature sensitivity and hysteresis in the sensing materials themselves Luo 2025↓Tactile Robotics: An OutlookShan Luo, Nathan F. Lepora, Wenzhen Yuan et al., IEEE Transactions on Robotics (accepted), 2025. Sensors also add cost and failure modes to collection rigs, and few policy architectures take tactile input natively Luo 2025↓Tactile Robotics: An OutlookShan Luo, Nathan F. Lepora, Wenzhen Yuan et al., IEEE Transactions on Robotics (accepted), 2025.

Four sensors anchor the accessible tier, and two of them are priced in three digits. GelSight Mini brings optical tactile sensing with high-resolution contact geometry to a commercial product and retails at $500 Luo 2025↓Tactile Robotics: An OutlookShan Luo, Nathan F. Lepora, Wenzhen Yuan et al., IEEE Transactions on Robotics (accepted), 2025. DIGIT, Meta FAIR's compact optical fingertip, is an open design with public files and a broad replication record, and retails at $350 Lambeta 2020↓DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand ManipulationMike Lambeta, Po-Wei Chou, Stephen Tian et al., IEEE RA-L, 2020 Luo 2025↓Tactile Robotics: An OutlookShan Luo, Nathan F. Lepora, Wenzhen Yuan et al., IEEE Transactions on Robotics (accepted), 2025. Its successor Digit 360 wraps the fingertip in 360° of optical coverage, resolves forces down to 1 mN across more than eight million taxels, and ships through GelSight under a manufacturing partnership announced in October 2024 Meta AI 2024↓Advancing embodied AI through progress in touch perception, dexterity, and human-robot interactionMeta AI, 2024. AnySkin takes the magnetic route: a soft, inexpensive skin whose learned policies transfer across replacement sensor instances without recalibration, the first uncalibrated tactile sensor to report that Bhirangi 2024↓AnySkin: Plug-and-play Skin Sensing for Robotic TouchRaunaq Bhirangi, Venkatesh Pattabiraman, Enes Erciyes et al., 2024.

Compute: closing the loop

NVIDIA built Jetson Thor specifically for this market. The T5000 module delivers 2,070 sparse FP4 TFLOPS with 128 GB of LPDDR5X at 273 GB/s on a 14-core Neoverse-V3AE CPU, in a 40-130 W envelope, rated 7.5x the performance of the AGX Orin it replaces; the T4000 trims that to 1,200 TFLOPS and 64 GB NVIDIA 2026↓Jetson Thor: Advanced AI for Physical RoboticsNVIDIA, 2026. Both are orderable through NVIDIA partners, and neither carries a published module price. The strongest signal that edge inference is the intended deployment path for humanoids is hardware: the 1X NEO's onboard computer is a Jetson Thor running at up to 2,070 FP4 TFLOPS 1X Technologies 2026↓NEO Home Robot1X Technologies, 2026.

VLA-Perf’s roofline model predicts π0 inference at 19.0 Hz on Jetson Thor, 32.2 Hz on RTX 4090, 61.7 Hz on A100, 162.5 Hz on H100 and 314.4 Hz on B100, without network latency Jiang 2026↓How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-PerfWenqi Jiang, Jason Clemons, Karu Sankaralingam et al., 2026. These are batch-one analytical predictions for the 2.7B π0 baseline, not measurements from running the policy on all five GPUs. The setup uses three 224×224 cameras, 32 language tokens (800 input tokens total), 14 action dimensions, chunk size 50 and 10 denoising steps, with BF16 or FP16 where BF16 is unsupported. The study excludes sensor and robot execution latency and distinguishes inference frequency from action execution frequency. Sixteen times the throughput across the range, and the edge module sits at the bottom of it.

In the B100 sweep, increasing denoising steps from 10 to 50 at chunk size 50 raises action-expert latency 5× and total VLA latency 2.15×. Increasing chunk size from 50 to 250 at 10 steps raises expert latency 40% and total latency 11% Jiang 2026↓How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-PerfWenqi Jiang, Jason Clemons, Karu Sankaralingam et al., 2026. These are different sensitivities, not proportional end-to-end throughput or zero chunk-size cost.

The GPU rows use the paper’s Table 10 configurations: RTX 4090 with 24 GB and 1,008 GB/s, A100 with 80 GB and 2,039 GB/s, H100 with 80 GB and 3,350 GB/s, and B100 with 192 GB and 8,000 GB/s Jiang 2026↓How Fast Can I Run My VLA? Demystifying VLA Inference Performance with VLA-PerfWenqi Jiang, Jason Clemons, Karu Sankaralingam et al., 2026. Memory bandwidth is in gigabytes per second, not the network table’s gigabits per second. These are the study’s configurations, not specifications for every SKU. Its modeled Thor bandwidth is 270 GB/s; that should not replace the separate vendor specification above.

This performance-study source does not establish a vendor price for any of these four GPU configurations. Their price fields remain “not disclosed” in this source scope, not because no manufacturer or seller publishes prices. No replacement price or market-wide absence claim is inferred. The latency budgets are developed in Real-Time Execution.

How to read these prices

Three rules. First, the as-of date: prices in this guide were observed between June and August 2026, except the LEAP Hand's $2,000, which is its build cost at 2023 publication, and hardware pricing moved repeatedly during the year. Second, currency: yuan figures are converted at July 2026 rates, and the table note keeps the original amount so the conversion is auditable. Third, honest gaps: read each missing value in its stated source scope. For the four GPU rows, a null price means this cited performance study does not establish a vendor price; it is not a claim about every seller. Tesla has released no Optimus 3 spec sheet, NVIDIA publishes no Thor module price, and Boston Dynamics does not sell Atlas. Those cells stay empty on purpose.