Robot comparison: engineering appendix ✨
The tables separate published component ratings, subjective reference-profile estimates and hypothetical future capabilities. The historical cutoffs are 31 December 2022 and 14 November 2026. Empty cells remain available for specifications that have not been established.
Batteries, motors, computing and sensing
| Component | 2022 | 2026 |
|---|---|---|
| Conventional high-power cell | Molicel P45B: 4.5 Ah, 3.6 V nominal, 45 A listed discharge current, 242 Wh/kg, maximum mass 69 g. Manufacturer timeline places launch in 2021. [1] | Molicel P50B: 5.0 Ah, 3.6 V nominal, 60 A listed discharge current, 260 Wh/kg, maximum mass 71 g. Publicly shown at CES 2024. [1] |
| Higher-energy high-power alternative | Amprius SiCore silicon-anode cell: company reports 360 Wh/kg and more than 3,000 W/kg, announced January 2025; 10C discharge without cooling and 15C with active cooling are separate stated operating conditions. [2] | |
| Specialist endurance-cell evidence | Amprius announced shipment of 450 Wh/kg cells in February 2022 for high-altitude aircraft. [3] | Amprius announced 500 Wh/kg SiCore500 cells in September 2026, measured at 1C; availability was expected in Q4. This announcement does not establish shipment by the cutoff. [3] |
| Geared brushless motor benchmark | DJI M3508 P19, documented in 2017: 365 g; 3 N·m continuous maximum torque at 469 rpm; approximately 147 W shaft output at that point. [4] | The same actuator class remains available. No quantitative era-wide improvement is assumed. [4] |
| Onboard compute example | Jetson AGX Orin 32GB: up to 200 INT8 TOPS at launch; configurable 15–40 W module power; production availability 3 August 2022. [5] | Orin NX Super: up to 157 sparse INT8 TOPS, 10–40 W modes. Larger alternative: Jetson Thor, available August 2025, up to 128 GB, 2,070 FP4 TFLOPS, 40–130 W. FP4 and INT8 figures measure different operations. [5] |
| Thermal imaging example | FLIR Hadron 640R, May 2022 datasheet: 640 × 512 thermal, 64 MP visible, approximately 56 g module mass, less than 1.8 W typical. [6] | FLIR Boson SX8, announced June 2026: 1280 × 1024 thermal, 60 Hz mode, less than 2.25 W, CAD-estimated 22.5 g without lens. Hadron includes a visible camera and different packaging, so these masses are not directly comparable. [6] |
Energy density, maximum current and power ratings apply under their specified test conditions. They cannot all be combined into an assumed operating point. The empty 2022 alternative cell means no matching dated high-power example has been selected here.
Complete-system reference profiles
Assume a compact ground vehicle roughly half a metre across, a prepared onboard controller, useful visible or thermal contrast, a flat floor and approximately 1–2 kg for the electrical battery subsystem. The intervals below are subjective 80% credible intervals conditional on that profile. Component evidence supports the ingredients; complete mass, power and thermal closure remains unfinished.
| Field | 2022 | 2026 |
|---|---|---|
| Starting mass limit | 10 kg | 10 kg |
| Total usable-energy limit | 50 MJ | 50 MJ |
| Assumed pack specific energy | 180–220 Wh/kg | 220–300 Wh/kg for the advanced silicon-anode option |
| Nominal battery electricity | 1.1 MJ; 80% interval 0.7–1.6 MJ | 1.4 MJ; 80% interval 0.8–2.2 MJ |
| Peak ground speed | 5 m/s; 80% interval 3–8 m/s | 5 m/s; 80% interval 3–8 m/s |
| Visual observation to motion command | 50 ms; 80% interval 25–120 ms | 30 ms; 80% interval 15–80 ms |
| Chosen structural material and protection specification | ||
| Verified complete mass budget | ||
| Verified continuous operating power and cooling | ||
| Counted active-machinery potential | ||
| Selected integrated weapon system | ||
| Measured performance from 100 m against the opponent pool | ||
| Working relative Elo | 0, by definition of the anchor |
The latency estimates end at an ordinary motion command and exclude physical response time. Empty resource and performance cells reflect unfinished specifications or missing evidence. Detailed autonomous weapon design and integration are omitted; the integrated weapon-system cells are deliberately blank.
For scale, using 80% of the central battery stock at an assumed constant 500 W gives 29.3 minutes in 2022 and 37.3 minutes in 2026. These are load-account examples; 500 W is an assumed load, and damage can end a contest much sooner.
Published weapon and capture systems
Named Ukrainian UGVs, gun stations and FPVs
| System and dated evidence | Published weapon and mobility | Published mass | Control described by the source |
|---|---|---|---|
| Droid TW 12.7 MoD adoption: 9 December 2024; frontline account: 25 April 2025. | Tracked UGV with M2 Browning 12.7 mm machine gun. Current manufacturer catalogue: maximum ground speed 12 km/h; stated travel range 50 km. [14] | Tablet remote control in MoD account. Manufacturer describes remote operation with automated detection and tracking. | |
| LIUT / Lyut MoD adoption and combat-test report: 18 March 2025. | Machine-gun UGV. Current manufacturer page lists 7.62 mm PKT / M240 and 28 km range on a charge. [15] | Manufacturer explicitly describes fully remote operation by a three-person crew. | |
| D21 Current manufacturer catalogue. | Electric all-wheel-drive UGV; listed remote battle-module options include an M2 machine-gun module. Maximum speed 10 km/h paved, 5 km/h off-road; listed range 16 km. [16] | Weight: 380 kg. Payload capacity: 200 kg. | Remote battle modules are expressly listed. Complete autonomous engagement authorization is unspecified. |
| Wolly 7.62 MoD report: 2 April 2025; development began 2022, initial trials in 2023. | Standalone remote weapon station for a 7.62 mm PKT machine gun; MoD reports frontline use. [17] | Reported station mass: 30 kg. | Remote operation with automatic detection and tracking. The release also gives up to 130 hours of operation.[a] |
| STING Manufacturer specifications; public interception footage by May 2025. | FPV interceptor. Manufacturer: maximum speed 280 km/h; endurance 6 minutes at maximum speed, up to 15 minutes at cruise. [18] | Maximum takeoff weight with payload: 4 ± 0.2 kg. Payload capacity: 500 g. | Radio-controlled FPV system; source describes operator training and pilot-dependent performance. |
| Shrike Official military news dispatch: 28 October 2023; Auterion/Skyfall update: 13 July 2026. | FPV strike drone. The 2026 announcement says deliveries with terminal guidance began that week. [19] | Terminal guidance is described as available. Swarming is described as a future software update. |
[a] Empty mass cells reflect absent all-up values in the examined primary sources. Current catalogue specifications may describe later configurations than the earliest dated field report. Wolly’s mass inclusion is unspecified, and its 130-hour figure describes operating endurance rather than an engagement-authorization policy. LIUT and D21 are separate products sharing one page; D21’s 380 kg must not be assigned to LIUT. ↩
| System and date | Published mechanism | Published autonomy and evidence |
|---|---|---|
| DeepMelt, November 2022 | Rotary-impact contact mechanism; NHRL three-pound class. | Organizer explicitly states fully autonomous control. Debut record: three wins, two losses, including one knockout. The contemporary revision is dated 16 November. [7] |
| Gonzalez et al., 2021 | Aerial net capture. | Onboard visual detection, pursuit and capture demonstrated experimentally. Favorable-background detections reported around 40 m; all-up aircraft mass is unreported in the paper. [8] |
| Switchblade 300, January–April 2022 | Loitering munition; carrier expended in the engagement. | Manufacturer describes autonomous navigation to designated coordinates, followed by operator confirmation and engagement. Pentagon briefing separately confirms delivery to Ukraine. Autonomous navigation supports one capability, while human authorization remains part of this published operating sequence. [9] |
| Chief-1, June 2025 | Aerial carrier with a shotgun-cartridge firing module. | Ukraine’s MoD reports authorization for service, automatic target acquisition, and manual and automatic firing modes. The short release does not specify the complete human-authorization or onboard-computation arrangement. [10] |
| Brave1-associated kinetic turret, May 2026 | Reusable kinetic counter-drone turret. | Ukraine’s MoD reports use by more than ten military units, automatic detection and tracking, and operator confirmation before engagement. [11] |
| Brave1-associated interceptors, June 2026 | Counter-drone interception. | Ukraine’s MoD reports combat testing and automation of 95% of the process. An operator selects a target and commands engagement before autonomous execution. The 95% figure concerns process automation, not interception probability. [12] |
These accounts have different evidential status: a competition organizer records matches, a research paper reports experiments, a manufacturer describes its product, and a defence ministry reports its systems. The public military descriptions give no common controlled benchmark from which to infer this arena’s win rates. Full runtime autonomy also requires all decisions and computation after the start to remain onboard.
Future-tier parameters
The ratings are retained working central estimates. Numerical future values below are engineering hypotheses defining the proposed packages.
| Relative Elo | Power, body and actuation hypothesis | Specific survival or restoration hypothesis | Selected weapon system |
|---|---|---|---|
| +400 | Broadly 2026 storage and electromagnetic actuation; stronger integration of perception and control. | Tenfold fewer mission-ending control/perception failures. A full +400 gain requires other causes of defeat to remain sufficiently small. | |
| +800 | Improved conversion and heat paths; illustrative 10 MJ operating stock within the total 50 MJ; electromagnetic actuators and isolated support paths. | Three or four independent substantial losses required to disable it versus one in the predecessor. | |
| +1200 | Laminated/cellular reconfigurable body, major actuators plus suitable small-scale actuation. | Useful geometry changes in 1–10 s; reconnect or replace selected components from carried inventory. | |
| +1600 | Protected 1–10 g working regions, with local sensing, computation, manipulation and accounted support. | Several localized losses required to remove useful overall capability; protection can remain shared. | |
| +2000 | Broad component fabrication from suitable material, integrated in a protected cellular body. | 600 s productive-machinery doubling; 95% fabrication efficiency for a specified positive free-energy increase. | |
| +2400 | Broader manufacturing library and fine integrated sensing/manipulation in grown or atomically precise material. | 120 s productive-machinery doubling; 99% fabrication efficiency under the same definition. | |
| Reservoir endpoint; Elo unassigned | General-purpose distributed fabrication, control, sensing and manipulation. | 60 s productive-machinery doubling; 99% fabrication efficiency, inherited endpoint assumptions. |
Weapon-system cells are deliberately blank. Optical mechanisms remain a speculative candidate requiring a complete source, optics, thermal and interaction account; no beam power or disabling range is assigned here. Exact future storage chemistry, motor specifications, component-library coverage and full resource budgets also remain open.
Energy, growth and rating calculations
Fixed usable-energy account
The 50 MJ allowance covers initially usable reserves, counted active-machinery potential and any additional usable energy obtained from the environment. Environmental contributions count cumulatively; the allowance is not replenished by spending earlier energy. Captured opponent inventory was already part of the contest’s initial accounting and introduces a transfer, rather than a new environmental source.
Designated pure structural material can supply structure and passive protection. Its exempt potential cannot power actions or provide a beneficial energy release. Reconfiguration and fabrication still require paid work. Ambient material can add mass; its availability, composition and processing costs remain physical constraints. The exact floor composition and productive starting fraction have not been selected.
50 MJ / 3.6 MJ per kWh = 13.8889 kWh
13,888.9 Wh / 10 kg = 1,388.89 Wh/kg
50 MJ as 450 Wh/kg cells = 30.86 kg of cells
50 MJ as 500 Wh/kg cells = 27.78 kg of cells
Fabrication efficiency and time
For a positive useful free-energy increase W stored in a manufactured product and efficiency η, input work is W/η and destroyed available work is W(1/η − 1). This describes an availability account. Physical heat also depends on enthalpy and entropy changes. It does not set motor or power-converter efficiency, and it does not determine the workload of making a neutral structure.
F(t) = F₀ × 2t/T
At t = 600 s: T = 600 s gives 2F₀; T = 120 s gives 32F₀.
Ratio = 16, before resource limits bind.
At the flag’s 300-second holding period, the factors are √2 ≈ 1.414 and 22.5 ≈ 5.657, a fourfold ratio. These curves assume unrestricted reinvestment into productive machinery. Allocating the same machinery to repairs or other products slows factory growth.
For an illustrative productive starting fraction of 1%, those unrestricted values are 2% and 32% of the initial body mass after ten minutes. The 1% is an illustration, not a selected design. Available matter, fabrication work, energy delivery, heat rejection and transport can all prevent the curve from being reached.
A single uncontested flag hold lasts five minutes, so a ten-minute growth comparison requires the contest to remain unresolved for longer than one holding period. A machine cannot assume that time while surrendering the objective.
Elo and conditional reliability models
Expected score = 1 / (1 + 10−Δ/400)
At Δ = 400: expected score = 10/11 = 90.909%.
Expected score = P(win) + ½ P(draw).
The double-loss rule also requires an explicit fitting convention. For this initial discussion, the 90% language concerns decisive adjacent-tier contests. A future tournament should report wins, losses, draws and double losses separately; a double loss is a failure of both designs and should not silently become a decisive result for either one.
For independent constant rates a and b of causing decisive loss of capability, P(new wins) = a/(a+b). A ten-to-one ratio gives 90.9%. Reducing one controller failure mode tenfold gives that result only if the total effective rate changes comparably.
If controller failures account for fraction f of baseline defeat hazard, a tenfold reduction in that component alone gives P(new wins) = 1/(2 − 0.9f) in a symmetric equal-remaining-capabilities model. At f = 0.90 this is 84.0%; at f = 0.99 it is 90.2%. The +400 hypothesis therefore requires unusually dominant controller failures or additional performance gains.
If the new machine needs n independent substantial losses to disable it, the old machine needs one, both cause losses at equal constant rates, and the new machine remains equally effective after partial losses, then P(new wins) = 1 − 2−n. At n = 3 this is 87.5%; at n = 4 it is 93.75%. Correlated losses or declining performance invalidate this simple calculation. It illustrates a scale of durability improvement that could support a rung.
If effective output instead falls proportionally with the number of surviving regions, the corresponding equal-initial-rate model gives 77.5% for three regions and 84.8% for four. Redundancy must preserve effective output, or another improvement must compensate, to support the higher figures.
These are analytical models. No physical simulation or match series currently measures the proposed ratings. Future calibration needs a diverse opponent pool, explicit objective adjudication, repeated conditions and inspection of nontransitive matchups. The working estimates can change when those comparisons expose a weaker or stronger transition.
Sources
- Molicel: P45B specifications, P50B specifications, manufacturer chronology, P50B at CES 2024.
- Amprius: 2025 annual report, filed in 2026.
- Amprius: 450 Wh/kg shipment, February 2022; SiCore500 announcement, September 2026.
- DJI: M3508 P19 guide; component specifications. Shaft power calculated as torque × angular speed.
- NVIDIA: AGX Orin 32GB, August 2022; Orin NX Super modes; Thor availability, August 2025.
- FLIR: Hadron 640R datasheet, revised May 2022; Boson SX8 specifications; June 2026 announcement.
- NHRL: DeepMelt revision dated 16 November 2022; competition record.
- Gonzalez et al.: Vision-based UAV Detection for Air-to-Air Neutralization, 2021.
- AeroVironment: January 2022 product announcement. US Department of Defense: 12 April 2022 briefing.
- Ukraine Ministry of Defence: Chief-1, 4 June 2025.
- Ukraine Ministry of Defence: AI-powered turret, 9 May 2026.
- Ukraine Ministry of Defence: Next-generation interceptors, 8 June 2026.
- Kaufmann et al.: Champion-level drone racing using deep reinforcement learning, 2023; University of Zurich account dating the races to June 2022.
- Ukraine MoD: Droid TW adoption, 9 December 2024; frontline account, 25 April 2025. DevDroid catalogue.
- Ukraine MoD: Lyut adoption, 18 March 2025. UFORCE land-platform catalogue, LIUT section.
- UFORCE land-platform catalogue, D21 section; weight and payload are separate manufacturer fields.
- Ukraine MoD: Wolly, 2 April 2025. DevDroid Wolly 7.62 catalogue.
- Wild Hornets STING manufacturer specifications. Manufacturer interception footage, dated by a 20 May 2025 report dating the footage to 19 May.
- ArmyInform Shrike dispatch, 28 October 2023; Auterion/Skyfall announcement, 13 July 2026.