Tool-first knee actuator screening
Estimate knee joint torque, power, thermal caution, and RFQ readiness before comparing quasi-direct drive, compact geared, or linear series-elastic actuator routes.
Route
single URL hybrid
Intent
tool + evidence
Updated
2026-06-05
Immediate result
Architecture review
Main CTA
Send knee RFQDefaults model a medium humanoid stair case. Adjust known values; invalid entries are clamped and recoverable.
Total robot mass without optional payload.
External payload that shifts knee demand during stance.
Approximate center-of-mass lever around the knee during the chosen scenario.
Advertised or measured actuator peak output torque.
Peak knee output speed at the joint, not motor shaft speed.
Percent of the cycle spent above roughly 60% of required peak torque.
Mass of one knee actuator module including motor, reducer/linkage, sensors, and structure.
Result band
Architecture review
The input set points to a mismatch. Treat this as an architecture review, not a routine quote request.
Required peak torque
284 N.m
Required peak power
2.8 kW
Torque margin
0.77x
Torque density
91.7 N.m/kg
Higher knee extension demand; treat thermal hold and brake fallback as early RFQ topics. Best for transparency and dynamic response when peak torque density is high enough.
Next step: Add stair height, cycle count, and bent-knee dwell time to the RFQ package.
90-360 N.m
Official Unitree public pages list G1 knee maximum torque at 90 N.m or 120 N.m by model, and H1 knee-class output around 360 N.m. Treat both as maximum/peak signals, not continuous ratings.
4 gates
Knee selection must clear torque, speed, thermal duty, and impact/backdrive validation before RFQ lock.
Test required
Public specs often state ultimate or peak torque; continuous low-speed knee duty needs winding temperature and cycle-test evidence.
ISO 10218-1/-2:2025
The 2025 ISO split keeps robot manufacturer duties and application/cell integration duties distinct.
Updated 2026-06-05. This pass focuses on content that could change an engineering or sourcing decision, not copy edits. Where public data is incomplete, the page now marks the missing proof instead of converting it into a stronger claim.
| Gap found | Before | Information added | Status |
|---|---|---|---|
| Peak torque range can be overgeneralized | A 120-360+ N.m range was useful but too easy to read as universal knee guidance. | Split public maximum knee torque into G1 90/120 N.m and H1 360 N.m model-specific examples, with date and peak/continuous boundary. | Strengthened with official Unitree sources |
| Safety boundary was too abstract | The page mentioned ISO 10218 without mapping it to what an actuator buyer should request. | Added robot/cell integration scope, drive-level safety-function boundary, and RFQ evidence items for brake, STO, and stop behavior. | Strengthened with ISO/IEC source context |
| Thermal claims had no pass/fail acceptance language | Thermal duty was flagged, but buyers still lacked a concrete validation package. | Added acceptance gates for RMS current, winding temperature, ambient/cooling setup, dwell time, and repeated-cycle traces. | Converted into supplier checklist |
| Human knee assist scenario lacked public benchmark limits | Wearable knee assist was listed, but the page did not separate it from humanoid knee actuation. | Marked exoskeleton/medical release data as application-specific and not interchangeable with humanoid peak torque benchmarks. | Boundary clarified; public evidence still limited |
Use these numbers as public reference anchors for actuator joints knee screening. They are not interchangeable with supplier acceptance data because maximum torque, event torque, and repeated continuous duty answer different questions.
90 N.m maximum knee torque
120 N.m maximum knee torque
around 360 N.m knee-class output
286 N.m, 15.5 rad/s, 1.5 kW event cited in jump context
| Step | Input | Output | Failure if skipped |
|---|---|---|---|
| Load case | Robot mass, payload, knee lever, gait scenario | Required knee peak torque with scenario multiplier | Sizing from robot mass alone without squat, stair, or stumble states |
| Power gate | Required torque and peak joint speed | Mechanical peak power at the knee | Meeting torque at a speed where the motor is already power-limited |
| Thermal gate | High-torque duty, cooling path, continuous torque evidence | Thermal caution band for repeated motion | Treating seconds-level peak torque as repeated gait capability |
| Architecture fit | QDD, compact gear, or linear/SEA layout | Torque-density target and validation emphasis | Choosing a gearbox ratio without reflected inertia and backdrive tests |
| RFQ evidence | Motion traces, acceptance criteria, safety fallback | Supplier clarification list and sample test plan | Asking only for a quote before freezing verification assumptions |
Public sources support screening ranges and architecture tradeoffs. They do not prove a supplier-specific continuous torque rating, cooling path, or lifetime claim. Those belong in the RFQ test plan.
| Source | Useful signal | Date | Confidence |
|---|---|---|---|
| Unitree G1 developer documentation | G1 public support data lists maximum knee joint torque at 90 N.m for G1 and 120 N.m for G1-EDU, with 0-165 degree knee motion range and 0.6 m calf + thigh length. | Official support page reviewed 2026-06-05 | Official model-specific maximum torque; not a continuous knee-duty rating |
| Unitree H1 developer documentation | H1 M107 joint motor and knee joint output are listed around 360 N.m; thigh/calf are 400 mm each. | Reviewed 2026-06-05 | Official support documentation |
| Variable reduction ratio humanoid knee research | Reports knee as the highest-demand lower-body axis in jumping; example events reach 286 N.m, 15.5 rad/s, and 1.5 kW. | arXiv HTML available 2025; reviewed 2026-06-05 | Research paper, not supplier catalog |
| CMU compact SEA for bipedal robots | Shows knee actuator requirements derived from scaled muscle torque and speed; early prototype knee rating was 12.6 N.m for swing-leg experiments. | CMU RI technical report 2011; reviewed 2026-06-05 | Research context with lower load target |
| RoMeLa linear SEA humanoid lower-body paper | THOR lower body used linear SEAs; knee pitch and hip pitch were serially actuated while ankle and other hip axes were parallelly actuated. | Humanoids 2014; reviewed 2026-06-05 | Architecture evidence |
| ISO 10218-2:2025 overview | ISO describes Part 2 as covering integration, commissioning, operation, maintenance, decommissioning, disposal, and reasonably foreseeable misuse for industrial robot applications and cells. | Published 2025; reviewed 2026-06-05 | Official standards scope; applies to robot application/cell safety, not actuator sizing |
| IEC functional safety FAQ | IEC lists IEC 61800-5-2 as the adjustable-speed electrical power drive systems functional safety standard; use it to frame drive safety functions such as torque-producing power removal. | IEC FAQ reviewed 2026-06-05 | Drive safety-function context; complete machine safety still needs system validation |
A candidate knee actuator should move from estimate to sample only after the supplier can answer these evidence requests. Mark unresolved items as pending instead of hiding them in a generic quote.
| Gate | Ask for | Accept when | Reject or hold when |
|---|---|---|---|
| Torque-speed curve | Output torque versus joint speed at bus voltage, not motor shaft speed. | Candidate clears required knee torque and speed at the same operating point. | Supplier provides only stall torque, no curve, or curve without voltage/cooling conditions. |
| Thermal repeatability | RMS current, winding or case temperature, ambient, cooling path, and repeated-cycle trace. | Temperature stabilizes below stated limit for the target duty and dwell time. | Peak torque demonstrated for seconds but no repeated stair, squat, or hold profile. |
| Impact and backdrive | No-power backdrive torque, reflected inertia, reducer/linkage shock load, bearing margin, and hard-stop assumptions. | Bench data covers stumble, landing, external disturbance, and emergency stop assumptions. | Only nominal walking data is available for a knee expected to absorb impacts. |
| Safety function boundary | Brake behavior, safe torque off or equivalent drive function, stop category strategy, and residual motion risk. | Actuator evidence is mapped into robot and cell risk assessment responsibilities. | Supplier implies an actuator module alone certifies the final robot application. |
| Risk | Probability | Impact | Mitigation |
|---|---|---|---|
| Continuous torque overstated from peak rating | High | High | Ask for RMS current, winding temperature, cooling boundary, and repeated-cycle test evidence. |
| Knee impact exceeds reducer or bearing margin | Medium | High | Define stumble, landing, hard-stop, and emergency-brake load cases before sample release. |
| Backdrive and impedance mismatch | Medium | Medium-High | Measure no-power backdrive torque, reflected inertia, and torque tracking under external disturbance. |
| Remote linkage geometry loses torque at critical knee angles | Medium | High | Run torque-speed checks at minimum, nominal, and worst mechanical advantage angles. |
| Safety claim made before cell integration | Medium | High | Separate actuator capability from robot and cell risk assessment under ISO 10218 and ISO 12100 logic. |
Inputs: 45 kg robot, 3 kg payload, 0.35 m lever, walking multiplier
Result: Roughly 165 N.m peak demand before detailed dynamic model
Action: QDD or compact geared joint can enter RFQ if thermal data is credible.
Inputs: 65 kg robot, 10 kg payload, 0.42 m lever, stairs multiplier
Result: Roughly 448 N.m first-pass peak demand
Action: Architecture review before supplier shortlist; do not rely on 300 N.m class headlines.
Inputs: Human-interactive joint, lower torque, high comfort sensitivity
Result: Torque tracking and impedance may matter more than peak torque density
Action: Series-elastic or low-backdrive path deserves parallel evaluation.
Inputs: 50 kg robot, 5 kg payload, 0.45 m lever, squat multiplier
Result: Above 460 N.m first-pass peak demand
Action: Request low-speed continuous torque, brake fallback, and thermal dwell tests.
There is no universal value. Official public examples include Unitree G1 at 90 N.m, G1-EDU at 120 N.m, and H1 knee-class output around 360 N.m, while research jump events can exceed 280 N.m. These are screening anchors, not continuous duty guarantees.
No. The knee also needs speed, power, thermal duty, impact margin, backdrive behavior, and control-loop evidence.
Early-stage users rarely have full inverse-dynamics traces. The multiplier is a conservative screening proxy that flags when detailed modeling is mandatory.
No. It is a first-pass RFQ triage tool. Use it to decide what evidence to request, then validate with gait simulation and bench tests.
Include calculator output, motion assumptions, control stack, sample quantity, and validation targets. The first review should close the missing evidence before architecture lock.
Email [email protected] or WhatsApp +86 18857971991.
