Robotic actuators: technology, innovation and manufacturing in India

Chapter 01

Executive brief

This paper evaluates how Armature AI Labs could develop compact, serviceable robot actuators in India. The first objective is a well-characterised V0 joint with a competitive mass and package, followed by improvements in sustained operation, heat rejection and precision. Matching catalogue mass alone is not an equivalent actuator: output support, electronics, sensing, connectors and cooling must fit inside the same comparison boundary.

The intended arm remains a general-purpose manipulator for grasping, pick-and-place and object positioning. Laboratory work, test tubes, fine assembly, biological experiments, soldering and PCB handling are additional development goals. They are not validated capabilities. A dexterous hand is a separate actuator, sensing and software programme that also adds wrist mass and contact complexity.

• Measure stationary holding and friction first. A joint may achieve a short peak torque yet be unsuitable for sustained gravity compensation or delicate contact.
• Prioritise winding-to-housing heat flow, current/torque calibration, dual-encoder diagnostics, repairable construction and a documented thermal estimator before pursuing exotic motor concepts.
• Treat OEM demonstrations as evidence of a robot task, not proof of undisclosed actuator life, thermal margins or supplier identity.
• Localise processes in stages: housings, shafts, harnesses, winding and assembly are different industrial problems from precision gears, encoders and sintered rare-earth magnets.
• Use the cost model to identify what quotations and yield data would change the decision. Its numbers are editable planning assumptions, not market prices.

How to review: engineering reviewers should challenge the load cases, thermal boundaries and test criteria; suppliers should correct specific capability or origin entries; business reviewers should replace cost assumptions with written quotations. Every referenced statement links to its source. Source IDs also resolve in the evidence appendix and spreadsheet. No supplier outreach or public publication was performed for this paper.

Chapter 02

Scope and evidence rules

Research cut-off: 25 October 2026. Public primary documentation, manufacturer disclosures and research publications underpin the evidence tables. A source date is retained where available; an undated webpage is identified as such. “Undisclosed” means the cited public material does not establish the fact, not that the manufacturer lacks the capability. Reported supplier capability is not a completed vendor audit.

A complete joint boundary includes motor, reduction, output bearings, encoders, driver, housing, harness exits, lubrication and any brake. Compare bare motors separately from complete joints. For cooled designs, also report the allocation of pumps, fans, radiators, coolant, hoses, power and structural heat sinks. Peak torque per kilogram, continuous torque per kilogram and installed arm mass answer different questions.

Platform mapping deliberately stops where public evidence stops. Patents describe possible embodiments; photographs suggest geometry; simulation files expose controller or modelling assumptions. None alone establishes a production supplier, gearbox topology or a guaranteed physical rating. A later generation is not automatically evidence about an earlier deployed fleet.

No paid supplier quotations, physical teardown, dynamometer measurements or component-origin audits were performed. Documented uncertainty is part of the result. The linked 24-entry comparison remains a separate specification reference; its exploratory arm load assumptions are not frozen requirements for this paper.

Chapter 03

The actuator as a complete system

A joint actuator is a controlled load path and a thermal system. A supplier may sell a bare motor, a gear unit, an integrated drive or a supported joint. Those are different procurement boundaries. The electronics, output bearing, encoder and brake can change mass and packaging as much as the motor choice. [E02] [E06] [E12] [E24]

Four different time claims. Warm-up time to a limit, sustainable operation at thermal equilibrium, duration/recovery of a short peak, and accumulated wear life are different questions. Running for eight hours at one light duty neither validates all loads nor establishes a thousands-of-hours service life. [E01] [E21] [E27]

Thermal limiting chain. Copper heating depends on current; magnetic losses rise with operating conditions including speed; the inverter contributes conduction/switching loss; gearbox/bearings/seals/lubricant dissipate mechanical losses. Windings, magnets, encoder, inverter, grease and touch surfaces can have different limits. [E01] [E02] [E11] [E14] [E17]

Holding is a separate operating point. At stationary output, mechanical output power is zero while energized torque can still generate copper and inverter heat. A running torque rating therefore does not automatically establish continuous zero-speed holding torque. Test multiple electrical rotor angles and the real mounting thermal path. [E01] [E11]

Precision and sensitivity. Backlash, elastic deflection, friction, cogging, energized ripple, encoder errors and thermal drift are distinct. Current-based torque estimation is not an actual output torque transducer; motor encoder resolution is not output positioning accuracy. [E04] [E06] [E13] [E18] [E26]

Installed mass and reflected inertia. For a simple reduction N = motor speed/output speed, rotor inertia referred to the output scales as N squared times rotor inertia; load inertia referred to the motor scales inversely. Distal mass also contributes gravity moment and link inertia. Comparisons should use a full robot model, not just module mass. [E03] [E02]

Architecture comparison

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

Architecture selection is an engineering assessment. Start from the joint duty and acceptable impedance, then size motor, transmission, bearing and thermal path together. High reduction can improve compact torque production but adds friction and reflected inertia; compliance can protect the mechanism but needs sensing and bandwidth analysis. [E03] [E04] [E05]

Chapter 04

Who leads, and what robots actually use

There is no evidence-based universal ranking. Commercial integration, sensitive torque control, precision gearing, miniature motors, field service and manufacturing scale reward different designs. The following specialist map describes evidence-backed niches, while the robot table distinguishes disclosed mechanisms from unknown joint details.

Specialists and application-specific leadership

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

For industrial mobile inspection, Spot and ANYmal offer more useful maturity evidence than isolated peak-torque records: Boston Dynamics reports a customer fleet above 1,500, while ANYbotics documents modular torque-controlled drives and a named maxon development/production partnership. This makes them informative references for integration and serviceability, not purchasable drop-in joint specifications. Individual lifetime and actuator mass remain qualification gaps. [R27] [R23] [R22]

For cost-sensitive dynamic platforms, Unitree exposes helpful joint-level limits and real packaging measures, including Go2 heat pipes and G1 dual encoders. OpenArm makes a named commercial motor-family mix inspectable, and explicitly uses a higher-reduction 4340 series where payload and packaging justify it. These are practical benchmarking routes; neither family maximum torque nor robot battery runtime establishes sustainable torque density. [R01] [R03] [R04] [R26]

For dexterous hands,1X's July 2026 forearm-motor/tendon disclosure is a strong architecture reference for reducing distal mass; Figure 03 instead supplies useful evidence about integrated tactile sensing and the industrialization of many actuator SKUs. Neither disclosure provides a common mass, cooling, accuracy and lifetime test from which to declare a universal winner. Their value to an India program is in identifying system-level experiments, not copying an unverified performance headline. [R21] [R11] [R13]

For high-load humanoids, compare generation-specific approaches: electric Atlas is distinct from its hydraulic predecessor; Apollo's disclosed linear architecture must be separated from current Apollo 2 claims; Digit 5's planned cycloidal drives are not Digit 4 hardware. Tesla's accessible recruiting disclosures establish rotary/linear development expertise but cannot substantiate a shipped joint BOM. Leadership in this category should be judged after comparable duty-cycle, field-reliability and service data exist. [R05] [R06] [R14] [R15] [R16] [R17] [R10]

Robot - generation - joint - actuator evidence

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

Scope is public primary evidence reviewed25 October 2026. A missing detail means it was not established in the reviewed evidence, not that the manufacturer has no internal data.

Manufacturer webpages, technical documentation, OEM engineering reports and a primary university thesis are separated by evidence_level. None of the manufacturer peaks or endurance figures in this matrix is independently verified here.

Joint-to-actuator rows deliberately stop at the source's disclosure level. Patents, external appearance, speculative supplier lists and simulation actuator limits are not used to infer installed production hardware.

Spot SDK supplemental documentation explicitly describes physical transmission mechanisms; those prose disclosures are retained. Its simulation-oriented torque/ratio table is not promoted to continuous hardware ratings. [R08]

A historical architecture remains historically useful without being attributed to a current generation:2018 ANYdrive SEA evidence is not automatically current ANYmal D/X evidence; Apollo 2023 linear disclosures do not closeApollo 2 joint-mapping gaps. [R23] [R14] [R15]

Commercial battery hours, robot fleet operating hours, actuator thermal equilibrium and component fatigue life answer different questions. Figure 02reported robot-hours, ANYmal C cycles andNEO hand cycles use different boundaries and cannot be ranked as equivalent lifetime data. [R12] [R24] [R21]

All Chinese platforms/actuators remain engineering benchmarks for the India-developed direction. This section is not a recommendation to buy finished Chinese actuators. Component origin disclosure and procurement decisions remain separate.

Chapter 05

Why heat, mass and precision interact

Thermal equilibrium and hours of operation

Engineering model: at zero speed, mechanical output power is zero even while the motor produces holding torque. Electrical and inverter losses can still be substantial. A simple thermal model is Cth dT/dt = Ploss - (T - Tamb)/Rth. For constant losses it approaches a steady rise of Ploss x Rth, with time constant Rth x Cth. More thermal mass delays heating; lower thermal resistance improves steady operation. These are different interventions. [E28]

These equations are an analytical starting point. Do not substitute DC bus current for phase RMS current, or phase-to-phase resistance for per-phase resistance without the winding connection and measurement convention. Motor torque constant conventions must match the current convention. Add inverter, iron, mechanical and gearbox losses for the tested operating point; temperature-dependent resistance creates feedback that a constant-loss model misses. [E01] [E11] [E17]

Illustrative calculation, not a component rating: with 30 W total heat and a 1.5 K/W system thermal resistance, steady temperature rise is 45 K. At 40 C ambient, the modelled node reaches 85 C. Halving thermal resistance lowers that node to 62.5 C. Doubling thermal capacitance only slows the rise. A winding hotspot may exceed that node, so a safe surface temperature does not establish winding or magnet safety.

Four distinct specifications are required: permissible continuous steady load; an intermittent peak envelope with recovery time; mission operating hours at a defined trajectory; and service life at defined loads and maintenance. A battery-runtime statement is not a joint-life statement. Eight hours of intermittent operation is not eight hours at peak torque. A short burn-in is not evidence for thousands of service hours. [E01] [E21] [E27]

Mass propagates upstream

Illustrative calculation: adding 0.5 kg at 0.6 m horizontal offset adds 2.943 Nm of gravity torque and 0.18 kg m2 of point-mass inertia about that axis. At 5 rad/s2 it adds 0.9 Nm of acceleration torque in this simplified model. Adding the same mass near the shoulder axis contributes much less. Thus distal driver, brake, wrist and hand mass can trigger a larger shoulder actuator, which then increases structure and support requirements. [E03]

A complete arm calculation must include link distributed inertia, motor rotor inertia reflected through the square of gear ratio, joint geometry, payload centre of mass, counterbalances, contact loads and cable forces. The simple example is a sensitivity demonstration, not an arm payload rating. A light motor with a heavy external bearing or cooling plate may lose its apparent system advantage. [E02] [E03]

Accuracy is a chain, not an encoder resolution

An encoder quantises angle; it does not remove mechanical compliance, transmission error, friction, thermal drift or mounting error. A motor-side encoder cannot directly observe all output-side deflection or backlash. An output encoder improves observability but still needs calibration. A physical torque sensor measures strain or displacement through a calibrated load path; current-based torque estimation includes uncertain friction, ratio efficiency and temperature effects. [E13] [E18]

For delicate contact, test both directions near zero speed. Report breakaway torque, hysteresis, minimum repeatable controllable output torque, noise, bias drift and closed-loop stability. A high command rate is not the measured torque bandwidth; timestamps, filtering, transport, controller update and mechanics all contribute latency. Tool-level force sensitivity also depends on kinematics and the tool contact geometry. [E03] [E04] [E18]

Speed, energy and wear

For illustration, 20 Nm at 40 rpm is about 83.8 W of mechanical output. A 2 kN linear force at 0.1 m/s is 200 W. These are not electrical-input ratings. High torque and maximum no-load speed must never be multiplied as if simultaneously available. Back-EMF, bus voltage, current limits, winding temperature and transmission speed limits shape the envelope. [E01]

Deceleration and lowering a load can return energy to the DC bus. Specify how much the battery or power supply can absorb, and design bus clamps, braking resistors or a coordinated motion strategy where necessary. A bench supply that sources current may not sink it. Power-loss tests must distinguish motor shutdown, controlled stop, brake engagement and stored elastic/gravitational energy. [E12]

Transmission wear, bearing preload, lubrication and sealing can change friction and accuracy before a joint fails completely. Endurance acceptance should include drift in lost motion, efficiency, temperature rise, vibration and noise, not merely continued rotation. An accelerated test is credible only when it preserves the relevant failure mechanism. [E14] [E27]

More thermal mass can delay a temperature rise without improving the eventual sustainable load. Improved steady cooling requires a better path to the environment. A heat pipe moves heat; it still needs an effective condenser/sink. The published heat-pipe experiment varies orientation and airflow, which are especially relevant when a robot joint changes posture. A phase-change insert should be rated by finite absorbed energy and its reset time, not advertised as indefinite cooling. The phase-change proposal and its test gate are engineering recommendations. [E28] [E30]

The limiting temperature can belong to the gearbox lubricant, encoder, drive electronics or exposed surface instead of the motor insulation. Kollmorgen explicitly discusses neighboring component limits in frameless robotic joints. Use the actual selected parts and their mounting conditions; do not reuse the example temperatures from that application note as universal design limits. A proposed component temperature register should list normal limit, protection threshold, measurement location, uncertainty and required margin for each part. [E02]

Chapter 06

Four application classes: provisional requirements

The classes prevent a low-friction wrist from being judged by the same priorities as an impact-loaded knee. For wrists and hands, moving mass and contact sensitivity are central. Shoulders need sustained gravity support; hips and knees need energy handling and impact tolerance; ankles combine support and contact transitions. Linear actuators must be evaluated through the linkage moment arm, including near-singular positions.

A proposed development screen is an eight-hour mission at stated ambient, with repeated motion/hold/idle segments and logged sensor data. Independently test stationary holding to equilibrium. A provisional near-steady screening criterion is less than 1 C total drift over the final 30 minutes at all measured winding, housing and driver nodes. Even this does not prove equilibrium: use a fitted thermal model and sufficient settling time, with predicted remaining rise below a declared margin. Stop at any predeclared component limit. Passing a particular mission does not qualify arbitrary trajectories.

Provisional application requirement matrix

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

Chapter 07

Materials, processes and manufacturing controls

The following materials are engineering candidates, not a released bill of materials. Selection depends on magnetic frequency, flux, temperature, fatigue, tolerance, contamination and corrosion requirements. A trade name or grade is insufficient without a process drawing and inspection plan. Shafts/axles are load-bearing components; robot joint axes are kinematic coordinates, not interchangeable material specifications.

Component-to-process matrix

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

A practical control plan follows the part through manufacture. Incoming material certificates do not establish the properties after machining, cutting, heat treatment, winding or bonding. Keep the material heat/lot, drawing revision, process recipe, inspection result, firmware and calibration version against the assembled serial number. This is a proposed traceability scheme.

Chapter 08

The Indian supplier and localisation network

This is a capability map for qualification, not an approved vendor list. It covers complete-joint developers, manufacturers, distributors, process partners and test facilities. Locations identify published business or facility evidence; they do not establish the origin of every part. Prices, MOQ and lead times remain quotation-dependent where public evidence is absent.

India can assemble a serious actuator supply chain now; a qualified complete joint is still a system project. Published capabilities cover motors, laminations, winding, machining, bearings, brakes, electronics and calibration. The unresolved work is model-specific qualification and integration, especially miniature reducers, bearing preload, encoder alignment and sustained thermal performance. This is an engineering synthesis, not a claim that any listed vendor can deliver an RS-equivalent joint today. [I05] [I09] [I11] [I17] [I25] [I26] [I30]

Do not describe India as having no NdFeB capability. ARCI has a government-confirmed sintered NdFeB pilot and PML now claims 2026 NdFeB production through Quantum Magnetics. Neither source establishes the requested motor arc grades, certified quantities, component origin, or delivery terms. Treat these as qualification leads. [I16] [I35]

Mineral sovereignty is not finished magnet supply. IREL oxide production, Bhopal metal mini-plants, strategic SmCo production, ARCI strip-cast-alloy pilot and the REPM industrial scheme are different stages. Obtain stage-by-stage country and facility declarations. Assembly, grinding, coating or magnetisation in India alone does not demonstrate Indian alloy or sintered-magnet manufacture. [I14] [I35] [I37] [I40]

Capacity discrepancy requires disclosure. The 25 March 2026 DAE release states 500 tonnes/year for the Vizag SmCo plant; the 2 April 2026 parliamentary release states 3 tonnes/year. This report does not reconcile them or use either value to promise supply. Request a plant-specific current production and commercial-allocation statement. [I38] [I39]

The 6,000-tonne REPM scheme is a future capacity target. The 13 August 2026 update describes bids and a two-year plant gestation period from award. It does not prove operating output or that a named bidder is an awardee. No scheme tonnage enters the near-term sourcing plan. [I37]

India invoices and India manufacture need separate columns. maxon, Renishaw and Mouser have official India channels, while SKF/Schaeffler/Timken/Mayr have local manufacturing evidence. Neither status proves the country of every SKU. Obtain supplier legal entity, ship-from location, country of manufacture and component/subassembly origin separately. [I06] [I17] [I18] [I19] [I23] [I25] [I28]

Test laboratories must match the measurand and drive technology. CPRI publishes an induction-motor scope; it cannot automatically be booked as a low-voltage BLDC joint qualification lab. Combine an agreed dynamometer/thermal fixture with reference torque measurement and calibrated dimensional inspection. Verify the current accreditation scope and measurement uncertainty. [I24] [I30] [I31] [I34]

Magnet supply chain: distinct industrial steps

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

Indian supplier capability and qualification register

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

For the existing arm options, NMotion and Kikobot remain complete-joint qualification candidates; Mechtex is a motor/manufacturing development route; maxon India is a configured component and drive route. These roles are not drop-in equivalence claims. Imported parts are allowed for the Indian V0 design, with origin disclosure. [I01] [I03] [I05] [I06]

All supplier quality/capability statements are first-party claims or official disclosures, not independently audited performance.

All test prescriptions, selection candidates and qualification gaps are proposed engineering actions. Source IDs support the underlying capability/material context, not validation of the proposed actuator.

Prices, MOQ and lead times are deliberately unresolved except the stated distributor ordering policy. Obtain domestic tax, freight, duties, currency and origin breakdown in any later RFQ.

Chapter 09

Innovation opportunities and priority experiments

Priorities below are engineering judgements based on benefit plausibility, evidence maturity, testability, development risk and fit with an Indian pilot supply chain. They are not numerical performance predictions. P1 means a strong first experiment; P2 needs additional integration work; P3 is a longer research bet. Compare at the same duty and include every auxiliary component.

Innovation assessment

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

The portfolio should be conditional. If winding temperature is limiting, improve the heat path before enlarging the motor. If gearbox friction sets the force floor, extra encoder bits alone will not solve sensitivity. If the duty consists mainly of long static holds, a brake or counterbalance may reduce heat more effectively than a cooling fan, but its mass, engagement shock and safety behaviour must be included. [E01] [E13] [E19]

Emerging magnet-lean or rare-earth-free motors merit exploration where supply-chain resilience is valuable. They should not be presented as a proven compact wrist replacement. Prototype at equal installed output, envelope, cooling and acoustic limits; compare controller complexity, torque ripple and efficiency over the mission, not a single favourable operating point. [E16] [E25]

Chapter 10

Manufacturing economics: what volume changes

The compared boundary is a motor, reduction, output support, driver, encoder, housing and harness assembled as one joint. Brake, external cooling and tool flange customisation are excluded and must be added if needed. Delivered-to-bench parts budgets include inbound logistics/import costs; the separate delivery allowance covers outbound finished-unit packing and freight. There is no separate blanket customs rate. GST, sales margin, finance, facilities overhead and upstream magnet/steel factories are excluded. The result is engineering batch cost, not a selling price.

Three routes illustrate different control over the supply chain. Domestic assembly minimises initial process development but retains component dependence. Progressive localisation introduces local winding, housings, shafts and selected machining. Deeper production develops the motor and precision transmission process; it does not imply every material or semiconductor is Indian, and does not include building a rare-earth refinery or magnet-sintering plant.

Parts-volume factors are assumed to be 1.00, 0.92 and 0.82 for 10, 100 and 1,000 good units respectively, for all three routes. Assembly labour is not discounted. Yield assumptions are shown with each batch result. These inputs are not negotiated discounts or measured factory yields.

Maintenance is a separate illustrative allowance of INR 1,500 per delivered joint for one planned service within an assumed first 1,000 operating hours, covering labour and consumables. This planning horizon is not a validated service interval or lifetime. Warranty reserve covers unscheduled defect support; it is not applied to the maintenance allowance. Replace both with actual service policy, failure data and labour/parts quotations. No end-user downtime cost is modelled.

Scrapped starts consume the full assumed parts and assembly budget; no salvage credit is included. Starts are a deterministic planning estimate, not a probability-based guarantee of batch yield. Final test and delivery are charged per good unit in this simplified model. It omits rework loops and test cost for failed units, so those should be added before investment decisions. Infrastructure is allocated fully to each isolated batch scenario; do not sum all three batch examples or interpret this as accounting depreciation.

The workbook separates engineering, tooling and test infrastructure inputs and shows starts, yield, recurring cost, reserve and fully allocated cost. Volume factors apply only to parts. A larger batch does not automatically improve yield or produce a discount: both require evidence. The example deliberately shows how local parts savings can be outweighed by development cost and process loss at early volumes.

Replace assumptions with quotations for the same frozen drawing, tolerance, coating, inspection, delivery quantity and origin declaration. Ask for one-off fixture and tool costs separately. Record replacement parts, calibration time, field service and supplier change control. For maintenance, budget labour and downtime as well as wear parts; the illustrative warranty reserve is not a reliability calculation.

Chapter 11

Qualification: from a promising joint to evidence

Every test begins with a calibrated measurement chain and a frozen hardware/firmware revision. Record uncertainty and synchronize torque, speed, position, current, bus voltage and temperatures. The acceptance criteria below are proposals; freeze them before testing and preserve failures as well as passes.

Test protocols and acceptance gates

Scroll horizontally for all fields. Citations open original sources; source locators are in the appendix.

A useful development fixture needs a reaction torque reference, speed/position measurement, controllable load, safe load restraint, temperature instrumentation and power measurement on the DC side. Add phase-current measurement when diagnosing copper and inverter loss. Steady holding tests need a torque fixture that does not require movement. Use resistance thermometry or embedded winding sensing to supplement surface temperatures. [E01] [E03] [I24]

Calibration and data practice: retain raw time-series files, units, sensor ranges, calibration certificates, sample timing, ambient and mounting photos, script version and analysis notebook. Compare repeat builds and report spread. An uncontrolled fan, changed grease fill or different mounting plate invalidates a thermal A/B comparison unless it is explicitly the variable under test.

Endurance sequence: measure baseline friction/lost motion and thermal rise; run a stated load/temperature spectrum; periodically repeat baseline measurements; inspect wear and lubrication at planned intervals. Report censored tests and failures. Establishing life needs enough samples, a failure definition and a statistical model; 100 hours of prototype operation establishes only that test exposure, not a service-life guarantee. [E27] [E14]

Chapter 12

A staged development and investment decision

A small parallel-gripper tool should precede a dexterous hand. The common interface needs a mechanically indexed datum, positive retention, defined electrical power and communications, tool identity and mass/centre-of-mass metadata. Repeatable attachment does not itself prove tool-position accuracy; check deflection and calibration after repeated changes. A wrist force/torque sensor, compliant fixture or remote-centre compliance may matter more for assembly than increasing peak motor torque.

For soldering, biological work and sensitive electronics, actuator development is only one layer. Task-specific fixturing, sensing, thermal isolation, contamination control, cable management and process verification need independent acceptance. Avoid converting a successful general pick-and-place demonstration into a claim of reliable biological or electrical work.

Recommended investment order: fund measurements and fixtures first; then one repairable V0 module and the most promising thermal-path change; then repeatability and endurance; finally production tooling. Defer expensive topology changes until measured loss and failure budgets show that conventional improvements cannot meet the target. The white paper supplies a decision framework, not a guarantee that a specific unbuilt actuator will succeed.