Studies

The Hand

The hand is where a humanoid either becomes useful or remains a demonstration. Three design families dominate, and each trades finger mass, force, and fidelity differently.

A human hand is commonly described as having around 27 degrees of freedom counting the wrist. No robot hand matches it fully. The research reference, the Shadow Dexterous Hand, reaches 24 joints, 20 of them actuated, by putting the motors in the forearm and pulling the fingers with tendons¹. It was the hand OpenAI used to solve a Rubik's cube one-handed with a policy trained entirely in simulation².

The opposite approach puts a motor in or next to every joint. The Allegro Hand and Carnegie Mellon's open LEAP Hand both use 16 direct-drive joints. LEAP's paper reports it can be built for about 2,000 USD in parts and assembled in roughly four hours³, which is why it has become the default hand in academic learning labs.

A third family is underactuated: fewer motors than joints, with linkages coupling each finger's curl. The Inspire RH56 drives 12 joints with 6 actuators⁴. It is compact and robust, and cannot move each finger joint independently.

Industry hands, as publicly stated

Commercial humanoid programmes publish little. Tesla's December 2023 Optimus Gen 2 video listed "11-DoF brand-new hands" with tactile sensing on all fingers⁵. In November 2024 Tesla engineer Milan Kovac posted that a new hand had 22 DoF in the hand and 3 in the wrist and forearm, and that the demo was teleoperated⁶. Figure's Helix post describes finger flexion and abduction inside a 35-DoF upper-body action space⁷. These are company statements, not peer-reviewed specifications.

Morph's position design target: begin with a LEAP-style direct-drive hand because it is cheap, reproducible and well studied, then test whether moving actuators into the forearm with tendons reduces finger mass enough to matter. See Milestone M1.

Touch

Cameras see the cup. Only skin knows it is slipping.

Two families of tactile sensor lead robot research. Vision-based sensors put a small camera behind a soft coated gel; when the gel presses on something, the camera sees its shape. GelSight established the approach and measures fine geometry and shear⁸. Meta's DIGIT shrank it into a low-cost fingertip⁹, and Digit 360 later made it fingertip-shaped and multimodal¹⁰.

Magnetic skins embed magnetic particles in an elastomer over a board of magnetometers. ReSkin is thin and replaceable¹¹; AnySkin focuses on swapping a worn skin without retraining¹². For a household hand that will wear out, swappability matters.

Open question for Morph: which sensor survives months of dishwater and laundry? We have found no published long-term household durability study. That gap is a candidate experiment (Open questions).

Muscle

A joint is a motor and a gearbox, and the gear ratio decides whether the robot can feel what it touches.

High gear ratios give a small motor large torque. The cost is backdrivability: push on the limb and the gearbox resists. Harmonic (strain-wave) drives reach high ratios in a single near-zero-backlash stage¹³; cycloidal drives tolerate shock well; planetary gearsets are backdrivable at low ratio.

MIT's Cheetah work made the case for quasi-direct drive: a large-diameter brushless motor with a low-ratio planetary stage, so joint torque can be estimated from motor current and impacts pass back into the motor rather than breaking teeth¹⁴. The low-cost Mini Cheetah showed this at small scale¹⁵. An older alternative, the series elastic actuator, inserts a spring and measures its deflection¹⁶.

The Body

Walking is now learned more than it is programmed — in simulation, thousands of robots at a time.

The turn came with GPU-resident simulators. Isaac Gym ran physics and learning on one GPU¹⁷, and Rudin et al. trained walking policies in minutes of wall-clock time with thousands of parallel robots¹⁸. Today Isaac Lab¹⁹ and MuJoCo's JAX port MJX with MuJoCo Playground²⁰ carry this forward.

Policies cross to hardware through sim-to-real techniques, the oldest being domain randomization²¹. Radosavovic et al. deployed a transformer trained entirely in simulation zero-shot on a real humanoid outdoors²². HumanPlus went beyond walking: a humanoid shadowing a person in real time from a single RGB camera, then imitating the skills it collected²³.

On hardware, the open research platforms are Unitree H1 and G1; Boston Dynamics replaced hydraulic Atlas with an electric Atlas in 2024, and in January 2026 unveiled a product version it describes as having 56 degrees of freedom, with 2026 deployments committed to Hyundai and Google DeepMind²⁴. Whole-body control — legs, torso and arms coordinated while manipulating — is where Figure's Helix 02 post says its newest model is heading, adding a 1 kHz "System 0" for balance and contact²⁵.

Morph's position design target: legs come last (M5). Household value is in the upper body; we would train locomotion in Isaac Lab or MJX only once the hand and arms work.

The Brain

A vision-language-action model is a language model that learned to answer with motion.

Google's RT-1 tokenized robot actions and trained a transformer on a large fleet dataset²⁶. RT-2 took the next step, co-fine-tuning a web-pretrained VLM to output actions as text tokens, and named the category "vision-language-action"²⁷.

The open lineage followed. Octo is a small transformer policy trained on 800k trajectories²⁸. OpenVLA is a 7B model trained on 970k demonstrations and reported a 16.5% absolute gain over the 55B RT-2-X across 29 tasks, with 7x fewer parameters²⁹; OpenVLA-OFT then showed how fine-tuning choices — action chunking, parallel decoding, continuous actions — change speed and success³⁰.

Physical Intelligence's π0 pairs the PaliGemma VLM with a flow-matching "action expert" producing continuous action chunks³¹. π0.5 co-trains on heterogeneous data and evaluates cleaning in homes it never saw³². Both are released as open weights in openpi³³, alongside the FAST action tokenizer³⁴.

Humanoid-specific models split thinking from acting. NVIDIA's GR00T N1 has 2.2B parameters, 1.34B in its Eagle-2 VLM, with a diffusion-transformer action module running at 120 Hz³⁵. Figure's Helix pairs a 7B VLM at 7–9 Hz with an 80M visuomotor policy at 200 Hz⁷. Google DeepMind's Gemini Robotics builds a VLA on Gemini 2.0³⁶.

Morph's position design target: we do not train a brain from scratch. We fine-tune open models — openpi (π0 / π0.5) first, OpenVLA-OFT as a comparison — on our own teleoperation data, and publish the results either way (M4).

Data

There is no internet of robot actions. Every trajectory has to be made.

The largest pooled source is Open X-Embodiment: data from 22 robots across 21 institutions, covering 527 skills³⁷. DROID took a different tack, one hardware rig in many places: 76k trajectories, about 350 hours, 564 scenes and 84 tasks³⁸. Humanoid-specific collections like AgiBot World are now appearing³⁹.

Teleoperation is how most of it is made. ALOHA made bimanual puppeteering cheap and introduced ACT⁴⁰; Mobile ALOHA put it on wheels and reports that, with 50 demonstrations per task, co-training with static ALOHA data raised success by up to 90%⁴¹. UMI removed the robot entirely: a handheld gripper with a camera, deployed later on robots⁴².

Two other sources scale further. Human video — Ego4D offers 3,670 hours of egocentric footage⁴³ and R3M showed it trains better visual encoders⁴⁴. And synthetic data — DexMimicGen multiplies a few demos into many in simulation⁴⁵, part of GR00T N1's data mixture³⁵.

Morph's position design target: build an open, LeRobot-compatible teleop rig (M3) and publish every household episode we record, consented and scrubbed of personal information. Tooling: LeRobot.

Safety & Evaluation

A home robot is judged by the one time it does the wrong thing.

Evaluation starts in simulation. LIBERO defines 130 language-conditioned tasks in four suites⁴⁶ and is now the default fine-tuning check for VLAs. SimplerEnv builds simulated replicas of real setups so results correlate with real-robot outcomes⁴⁷. CALVIN tests chains of instructions⁴⁸, and BEHAVIOR-1K grounds 1,000 household activities in what people say they want robots to do⁴⁹.

Physical safety has standards. ISO 13482 covers personal care robots⁵⁰; ISO/TS 15066 gives limits for contact between collaborative industrial robots and people⁵¹; it does not formally apply to home robots, and ISO 10218-1:2025 explicitly excludes household consumer products from its scope — a gap in the standards a home humanoid falls into. Semantic safety is newer: DeepMind's ASIMOV benchmark asks whether a model understands which actions are unsafe in context⁵².

Morph's position design target: no unsupervised operation in a home. Every pilot (M6) runs with a person present and a physical stop, and we publish failures alongside successes.

References

1. Shadow Robot Company — Dexterous Hand Series. shadowrobot.com
2. OpenAI et al., "Solving Rubik's Cube with a Robot Hand", 2019. arXiv:1910.07113
3. Shaw, Agarwal, Pathak, "LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning", RSS 2023. arXiv:2309.06440
4. Inspire Robots — dexterous hands. inspire-robots.com
5. Ars Technica, "Tesla unveils its latest humanoid robot, Optimus Gen 2", Dec 2023. arstechnica.com
6. Cybernews, reporting M. Kovac's post on the Optimus hand, Nov 2024. cybernews.com
7. Figure AI, "Helix: A Vision-Language-Action Model for Generalist Humanoid Control", 2025. figure.ai
8. Yuan, Dong, Adelson, "GelSight: High-Resolution Robot Tactile Sensors…", Sensors 2017. mdpi.com
9. Lambeta et al., "DIGIT", 2020. arXiv:2005.14679
10. Lambeta et al., "Digitizing Touch with an Artificial Multimodal Fingertip", 2024. arXiv:2411.02479
11. Bhirangi et al., "ReSkin", 2021. arXiv:2111.00071
12. Bhirangi et al., "AnySkin", 2024. arXiv:2409.08276
13. Harmonic Drive LLC — strain wave gearing technology. harmonicdrive.net
14. Wensing et al., "Proprioceptive Actuator Design in the MIT Cheetah", IEEE T-RO 2017. ieeexplore
15. Katz, Di Carlo, Kim, "Mini Cheetah", ICRA 2019. ieeexplore
16. Pratt, Williamson, "Series Elastic Actuators", IROS 1995. ieeexplore
17. Makoviychuk et al., "Isaac Gym", 2021. arXiv:2108.10470
18. Rudin et al., "Learning to Walk in Minutes…", CoRL 2021. arXiv:2109.11978
19. Isaac Lab. github.com/isaac-sim/IsaacLab
20. Zakka et al., "MuJoCo Playground", 2025. arXiv:2502.08844; MJX docs mujoco.readthedocs.io
21. Tobin et al., "Domain Randomization…", 2017. arXiv:1703.06907
22. Radosavovic et al., "Real-World Humanoid Locomotion with Reinforcement Learning", 2023. arXiv:2303.03381
23. Fu et al., "HumanPlus", 2024. arXiv:2406.10454
24. Boston Dynamics, "An Electric New Era for Atlas", 2024. bostondynamics.com; product unveiling, Jan 2026, bostondynamics.com
25. Figure AI, "Introducing Helix 02: Full-Body Autonomy". figure.ai
26. Brohan et al., "RT-1", 2022. arXiv:2212.06817
27. Brohan et al., "RT-2", 2023. arXiv:2307.15818
28. Octo Model Team, "Octo", 2024. arXiv:2405.12213
29. Kim et al., "OpenVLA", 2024. arXiv:2406.09246
30. Kim, Finn, Liang, "Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success", 2025. arXiv:2502.19645
31. Black et al., "π0", 2024. arXiv:2410.24164
32. Physical Intelligence, "π0.5", 2025. arXiv:2504.16054
33. openpi. github.com/Physical-Intelligence/openpi
34. Pertsch et al., "FAST", 2025. arXiv:2501.09747
35. NVIDIA, "GR00T N1", 2025. arXiv:2503.14734
36. Gemini Robotics Team, "Gemini Robotics", 2025. arXiv:2503.20020
37. Open X-Embodiment Collaboration, 2023. arXiv:2310.08864
38. Khazatsky et al., "DROID", 2024. arXiv:2403.12945
39. AgiBot World Colosseo, 2025. arXiv:2503.06669
40. Zhao et al., "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware" (ALOHA / ACT), 2023. arXiv:2304.13705
41. Fu, Zhao, Finn, "Mobile ALOHA", 2024. arXiv:2401.02117
42. Chi et al., "Universal Manipulation Interface", 2024. arXiv:2402.10329
43. Grauman et al., "Ego4D", 2021. arXiv:2110.07058
44. Nair et al., "R3M", 2022. arXiv:2203.12601
45. Jiang et al., "DexMimicGen", 2024. arXiv:2410.24185
46. Liu et al., "LIBERO", 2023. arXiv:2306.03310
47. Li et al., "Evaluating Real-World Robot Manipulation Policies in Simulation" (SimplerEnv), 2024. arXiv:2405.05941
48. Mees et al., "CALVIN", 2021. arXiv:2112.03227
49. Li et al., "BEHAVIOR-1K", 2024. arXiv:2403.09227
50. ISO 13482:2014, Robots and robotic devices — Safety requirements for personal care robots. iso.org
51. ISO/TS 15066:2016, Collaborative robots. iso.org
52. Sermanet et al., "Generating Robot Constitutions & Benchmarks for Semantic Safety", 2025. arXiv:2503.08663