DISCUSSIONSTARTED BY WBH
Dexterous hands: how many degrees of freedom do you actually need?
Hardware & ownershipGalbot ET1Galbot G1VegaWBH Editorial ·
TOPIC HUB · /topics/manipulation
Manipulation
Humanoid manipulation: grippers and hands, teleoperated demonstrations, ACT and diffusion policies, MoveIt, and robots with published arm payloads.
47 ROBOTS · 7 UPDATES
Introduction
Manipulation is reaching, grasping and moving objects with a humanoid's arms and hands. The robots in this database range from parallel grippers to multi-fingered hands, and their published arm payloads differ widely; each record below lists its end effector and payload with sources.
Much current work learns manipulation from demonstrations. An operator teleoperates the robot, each demonstration is recorded as a dataset, and a policy learns to imitate it; LeRobot provides this record-and-train loop for real robots, including humanoids [1]. Two policy families are common starting points: ACT, which predicts short sequences of actions with a transformer [2], and Diffusion Policy, which generates actions by iterative denoising [3].
Classical pipelines remain important where tasks are well defined: MoveIt 2 provides motion planning, kinematics and manipulation for robots on ROS 2 [4]. Simulated benchmarks help compare both approaches; Isaac Lab includes pick-and-place tasks for humanoids such as the Unitree G1 [5].
CONCEPTS · 4
Key concepts
- Teleoperated demonstrations
- Recordings of a person controlling the robot, used as training data for imitation learning [1].
- Action chunking
- Predicting a sequence of future actions at once rather than one step at a time, as in ACT (Action Chunking with Transformers) [2].
- Diffusion policy
- A visuomotor policy represented as a conditional denoising diffusion process over actions [3].
- Motion planning
- Computing a collision-free arm trajectory to a goal; MoveIt 2 is the standard planner on ROS 2 [4].
KNOWLEDGE GRAPH · 47 ROBOTS
Robots and Manipulation
| ROBOT | END EFFECTOR | ARM PAYLOAD |
|---|---|---|
| 4NE1 Gen 3.5NEURA | Integrated high-dexterity hands; exchangeable forearms | Not published |
| 4NE1 MiniNEURA | Not published | Not published |
| AGIBOT A2 (A2 Ultra / A2 Lite)AGIBOT | Not published | ≈2 kg (A2 Ultra / A2 Lite) |
| AGIBOT A2-WAGIBOT | Not published | 5 kg per arm |
| AGIBOT A3AGIBOT | Not published | Not published |
| AGIBOT A3 UltraAGIBOT | Not published | 5 kg per arm |
| AGIBOT Genie G1AGIBOT | 6-DoF hand, replaceable with two-finger gripper | 3 kg continuous, one arm |
| AGIBOT Genie G2AGIBOT | Not published | 5 kg per 7-DoF arm, joint torque sensors |
| Agility Digit 5Agility | Swappable grippers (ISO-standard mounting flanges) | Not published |
| Apollo 2Apptronik | Not published | Not published |
| Astribot S1Astribot | Not published | 5 kg per arm (horizontal reach) |
| Booster K1Booster | Not published | Not published |
| Booster T2Booster | Optional; dexterous hands or grippers | Max 10 kg (posture dependent) |
| Boston Dynamics AtlasBoston Dynamics | Grippers, tactile fingers and palm | 20 kg (44 lbs) |
| EngineAI PM01EngineAI | Not published | Not published |
| EngineAI T800EngineAI | Not published | Not published |
| Figure 03Figure AI | Five-finger-style hands with fingertip tactile sensors (Figure claims 3 g sensitivity) and palm cameras | Not published |
| Fourier GR-3Fourier | 6-DOF base hand, optional 12-DOF hand | ≈3 kg |
| Galbot ET1Galbot | Not published | Not published |
| Galbot G1Galbot | Swappable dexterous hands, suction cups, grippers | 5 kg single arm, 10 kg total |
| HMND 01 Alpha BipedalHumanoid | 12-DOF 5-finger hand or 1-DOF parallel gripper | Not published |
| HMND 01 Alpha WheeledHumanoid | 12-DOF 5-finger hand or 1-DOF parallel gripper | Not published |
| iCubIstituto Italiano di Tecnologia | Not published | Not published |
| IRON (next-generation)XPENG (XPeng Inc.) - robotics business | Not published | Not published |
| LimX LunaLimX Dynamics | Not published | 3 kg |
| LimX OliLimX Dynamics | 5-finger hand (6 DoF), 2-finger gripper (1 DoF), 3-finger gripper (4 DoF) | 3 kg (single arm) |
| MenteeBot V3Mentee | Not published | Not published |
| NEO1X | Five-finger tendon-driven hands, force-controlled and backdrivable, with tactile/shear skin (per 1X hands post) | 18 lb (as published) |
| Noetix BumiNoetix | Not published | Not published |
| Noetix E1Noetix | Not published | Not published |
| Noetix N2Noetix | Not published | 5 kg (arm max load) |
| PAL Robotics TALOSPAL | 3-finger gripper (modular, customizable) | 6 kg (arm fully extended) |
| Persona humanoidPersona AI | Not published | Not published |
| PhantomFoundation | Not published | Not published |
| Pollen Robotics Reachy 2Pollen | Parallel gripper (interchangeable tips), 10N nominal / 50N peak | ~3 kg per arm |
| ROBOTERA L7ROBOTERA | ROBOTERA XHAND 1 dexterous hand (fully direct-driven, 12 active DoF) | 20 kg (dual-arm) |
| ROBOTERA M7ROBOTERA | 12-DoF direct-drive dexterous hand | Not published |
| ROBOTERA Q5ROBOTERA | XHAND 1 Lite dexterous hand | 10 kg (dual-arm maximum) |
| SproutFauna | Integrated 1 DoF grippers with rubberized interior | Not published |
| ToroboTokyo | Optional: hand (Torobo Hand) or in-house gripper | 7 kg per arm (worst posture) |
| Unitree G1Unitree | Optional Dex3-1 three-finger hand (G1 EDU) | ~2 kg (G1) · ~3 kg (G1 EDU) |
| Unitree G1-CompUnitree | Not published | Not published |
| Unitree G1-DUnitree | Not published | Approx. 3kg |
| Unitree H2Unitree | not published (H2) · Dexterous hand (multiple model options) (H2 EDU) · Optional 5-finger tactile hand, 22 active DOF (H2 Plus) | ~15 kg peak |
| Unitree R1Unitree | Optional dexterous hand (R1 EDU) | ~2 kg |
| VegaDexmate | Ordered separately: Dexmate 6-DoF hand, DexGripper D or S, or own tool | 5 kg (10+ lbs) per arm |
| Walker S2UBTECH | Not published | Not published |
Values come from each robot's record, where every fact links to its source.
INTELLIGENCE
Latest updates
- Unitree publishes 40 more G1 Dex1 manipulation datasets
- Figure introduces Helix 2.5, claims zero-shot whole-body tasks in 30 unseen homes
- Unitree publishes 44 G1_WBT household-task datasets
- Unitree open-sources UnifoLM-WLA and releases the UnifoLM-ER models
- 1X details 25-DoF tendon-driven hands that will ship on every NEO
- Figure 03 begins parts-sequencing work at BMW Plant Spartanburg
- Isaac Lab 2.3 adds a G1 loco-manipulation environment
RESEARCH
Research
0 PAPERS
No research has been published on this topic yet.
Papers are reviewed in weekly research digests before anything is published.
COMMUNITY · 1 THREAD
Community discussions
Community posts are written by members and are not checked by WBH. Facts on robot pages come from cited sources; report a data error to have it corrected.
LEARN