mars-autonomy

Mars autonomy

capability Hard import robotics
TRL Mars
Energy intensity
Required by
0
Requires
2

Decision-making capability for Mars-side robots, vehicles, and habitat systems. Spans onboard machine perception, planning, manipulation, and intervention decision under 8–48-minute Earth latency. Architecture: end-to-end vision-language-action (VLA) models on-device (NVIDIA Jetson AGX Thor / Tesla HW-class compute); Mars-side supervisor crew for non-routine authorization; ground-side review + replanning over multi-sol cycles. Tesla FSD + Optimus + Boston Dynamics + Figure operational systems provide the Earth-validated heritage.

Last reviewed: 2026-06-09

Governing equations

Permitted decision latency must be many orders of magnitude faster than Earth round-trip. Onboard: ms-s. Earth-supervised: hour-sols. [1]

Onboard inference throughput for vision-language-action models. NVIDIA Jetson AGX Thor: 200 TOPS at 60-100 W. Sufficient for end-to-end manipulation + navigation models on-device. [2]

Probability of requiring Earth intervention. Reduced by reliable onboard autonomy (P_auto-recover) — modern VLA + classical planners achieve > 99 % auto-recovery for routine tasks. [1]

Earth-Mars supervised operations loop. Classical Mars rover ops: 1-2 sols per plan-execute-review cycle. Autonomous: minutes-hours. [1]

Key constants & quantities

Symbol Value Units Conditions Description
t_round-trip,min 480 s (8 min — Mars opposition) Minimum Earth-Mars round-trip light time. Even at closest approach, real-time teleoperation is impossible.[3]
t_round-trip,max 2,880 s (48 min — solar conjunction) Maximum Earth-Mars round-trip light time. Near solar conjunction, ground commands may not arrive in time for routine operations.[3]
TOPS_onboard,humanoid 100–300 TOPS (Trillion Operations Per Second) Onboard inference compute for modern humanoid robots. Sufficient for transformer-based VLA models at sub-second latency.[2]
P_compute,inference 60 ±10 W W (NVIDIA Jetson AGX Thor) Onboard inference compute power. Significant fraction of robot energy budget; partly offsetable by hardware acceleration.[2]
d_AutoNav,Perseverance 100 ±50 m/sol m / sol Perseverance AutoNav daily autonomous traversal. Established proof of Mars-flight-validated autonomy.[1]
t_supervisor-cycle,Mars 60–600 s (Mars-side supervisor review) Time for Mars-side crew to authorize non-routine robot action. Real-time + minutes vs Earth-supervised sols.[1]
τ_software-update 26 months (Mars window) Earth-to-Mars software update cadence tied to 26-month launch windows. Software shipped at next window must serve full mission duration.[3]
rate_VLA-model-Q1-2026 80 % task success on Mars-analog manipulation Approximate end-to-end VLA model success rate on Mars-analog tasks (Figure 02 BMW deployment data; Optimus mode test results, Q1 2026).[2]

Operating envelope

ParameterRangeUnitsSource
Round-trip latency 480 – 2880 s [3]
Onboard compute 100 – 300 TOPS [2]
Task autonomy success rate 80 – 99 % on validated tasks [2]
Mars-supervisor decision latency 60 – 600 s [1]
Earth-side review cycle 21600 – 86400 s (6-24 h per cycle) [1]

Mass balance

Basis: 4-crew Mars base + 4 humanoid robots + 2 autonomous rovers, 1 year operations

Inputs

Onboard compute hardware (one-time) 12 kg (across all robots + vehicles) [2]
Software updates (data ingress) 100 GB/year [2]
Telemetry / logs (data egress) 4,000 GB/year [1]
Inference electrical (across fleet) 4,400 kWh/year [2]
  • Onboard compute hardware (one-time): Jetson + Tesla HW + redundant compute. Replacement at end-of-life.
  • Software updates (data ingress): Model updates + parameters + procedure updates. Compressed for Earth-Mars transmission.
  • Telemetry / logs (data egress): Robot + system telemetry to Earth for review + learning. Bandwidth-budget item for laser link.
  • Inference electrical (across fleet): ~ 0.5 kW continuous × 6 robots × half-duty cycle.

Outputs

Autonomous decisions executed 50,000,000 per year (across fleet) [2]
Mars-supervisor interventions required 50,000 per year (~ 0.1 % of routine) [1]
Earth-side review-cycle data products 5,000 per year [1]
  • Autonomous decisions executed: ~ 100 decisions/min × 6 robots × 8 h/sol × 220 sols.
TRL · Earth
8/ 9
TRL · Mars
5/ 9
Earth autonomy: TRL 8-9 for narrow domains (FSD highway, warehouse manipulation, ISS Astrobee). TRL 6-7 for general-purpose humanoid (Optimus / Figure / Apollo / Digit in industrial pilots). Mars autonomy: TRL 5 for routine rover operations (Perseverance AutoNav established); TRL 4 for crew-equivalent humanoid autonomy at Mars conditions. Software updates over 26-month windows are the bottleneck.[2]
Energy budget
0.5 kWhe / kWh / inference-hour (typical) [2]

Onboard inference + robot control. Negligible per individual decision; cumulative across fleet of 6 robots × continuous operation: ~ 4.4 MWh/year.

Variants & trade-offs

End-to-end VLA model (Tesla / Figure / Physical Intelligence)

[2]

Single transformer-based model directly mapping (vision + language + state) → joint actions. Trained on millions of hours of tele-operation + simulation + reinforcement learning. Tesla Optimus end-to-end, Figure 02 with Helix, Physical Intelligence π0 / π0.5.

Parameters
1000000000–10000000000 (1-10 B)
Inference latency
0.01–0.1 s
Task success rate
70–95 %
Stack lifetime
0–0 capability (Mars-window refreshable)
Materials: Multi-camera vision input · Proprioceptive sensor fusion · NVIDIA Jetson / Tesla HW custom ASIC · Pre-trained foundation model weights
  • Highest task generality (millions of distinct tasks possible)
  • Continual improvement via fleet learning
  • Direct port of Earth-validated industrial heritage
  • Robust to novel situations (within training distribution)
  • Failure modes hard to predict + debug
  • Generalization edge cases (Mars novelty out of distribution)
  • Updates require Earth-window cycle
  • Mars-radiation tolerance unproven for ML inference accelerators

Classical planning + perception (Perseverance AutoNav heritage)

[1]

Explicit terrain mapping + planning + obstacle avoidance. Mars rover heritage since Sojourner. Predictable, debuggable, conservative. AutoNav on Perseverance: real flight-validated.

Decision latency
2–30 s per planning cycle
Daily autonomous distance
50–200 m/sol
Task domain
0–0 Narrow (navigation, sample collection)
Stack lifetime
50000–200000 h operational
Materials: RAD750 + Mars-rated FPGA compute · Mars-validated software stack (VxWorks RTOS) · Multi-camera + LiDAR perception · NASA-grade verification toolchain
  • Highest TRL on Mars (TRL 9)
  • Predictable + debuggable behavior
  • NASA-grade reliability certification
  • Decades of operational heritage
  • Narrow task generality (navigation + obstacle avoidance only)
  • Slower than VLA models
  • No general manipulation or judgment beyond hand-coded heuristics
  • Difficult to extend to new tasks without ground-software cycle

Hybrid VLA + classical planning + Mars-supervisor crew

[1]

VLA model for routine tasks; classical planner as safety oversight + fallback; Mars-side crew for non-routine authorization. Combines best of all worlds. Likely Mars-mission architecture.

Autonomous handling
95–99.9 %
Mars-supervisor handling
0.1–5 %
Earth-supervisor handling
0–0.1 %
Stack lifetime
0–0 capability (hybrid)
Materials: VLA model + classical planner co-deployment · Mars-side mission supervisor crew interface · Earth-side review + planning ground software
  • Combined task generality + safety oversight
  • Mars-side crew can authorize edge-case decisions in seconds
  • Earth-side review for learning + improvement
  • Aligned with NASA flight-software practices
  • Higher system complexity
  • Mars-supervisor workload affects crew time budget
  • Interface design between VLA + classical + crew must be tight

Failure modes

Mode Cause Detection Mitigation
VLA model out-of-distribution failure[2] Robot encounters task or environment outside training distribution; model produces incorrect or unsafe action. Anomaly detection (confidence scoring); Mars-supervisor crew alert; safety-monitor classical planner. Conservative classical planner as safety oversight; Mars-supervisor crew alert on confidence drop; safe-mode default action; periodic Earth-side review for distribution updates.
Compute hardware SEU[4] Mars-surface GCR + SPE causes single-event upset in inference accelerator; rare incorrect inference. Inference cross-check (TMR); model output anomaly detection. Mars-radiation-rated compute (where possible); ECC memory; redundant inference compute; safe-mode + reset on detected upset.
Software update window slip[3] Earth-side software issue delays update; missed 26-month launch window means another Mars cycle without fix. Pre-launch software validation + update planning. Conservative software development cycles; pre-shipped contingency patches; over-the-air update over laser link if mission-critical (slow but possible).
Mars-supervisor crew incapacitation[3] Crew member sick / EVA accident / sleeping; no Mars-side authorization for non-routine decisions. Crew-rotation tracking; emergency protocol. Multiple crew members trained as supervisor; conservative auto-default actions; Earth-side authorization for safe-mode operations.
Adversarial input / sensor spoofing[1] Sensor failure or environmental anomaly (mirror reflection, low-contrast terrain, dust occlusion) confuses perception. Cross-sensor consistency check; classical-planner sanity-check on VLA output. Multi-modal sensor fusion (vision + LiDAR + IMU); periodic sensor health check; conservative behavior in low-confidence regions.
Cascading task failure[1] Single robot task failure cascades into multi-robot coordination collapse (e.g. 1 robot stuck blocking another). Coordination protocol monitoring; task completion tracking. Designed-for-degraded-mode coordination (one robot fault should not stop fleet); periodic full-fleet status review; Mars-supervisor manual intervention.
Mars-Earth comms outage during autonomy decision[3] Solar conjunction or laser-link fault prevents Earth-side review during critical decision. Comms link status alarm. Pre-authorized Mars-supervisor authority for conjunction periods; conservative autonomous safe-mode defaults; pre-stored Earth contingency procedures.

Mars adjustments

Earth latency makes real-time supervision impossible[3]

Impact: 8-48 minute round-trip means no teleoperation, no real-time correction, no Earth-side oversight at second-scale. Architecture must be inherently autonomous + Mars-side-supervised.

Mitigation: High onboard autonomy (VLA models); Mars-side supervisor crew for non-routine; pre-cached Earth contingency procedures; multi-sol Earth review cycles.

Software updates tied to 26-month windows[1]

Impact: Software shipped at next Mars window must serve full mission duration. Earth-side update over laser link is slow + bandwidth-limited (especially during conjunction).

Mitigation: Conservative software development; pre-shipped contingency patches; over-the-air capability for emergency only; on-Mars-base local model refinement using onboard compute.

Mars-radiation tolerance of inference compute[4]

Impact: NVIDIA Jetson + Tesla HW silicon designed for Earth ground use. Mars surface GCR + SPE flux 10× Earth LEO; cumulative TID + SEU rate higher.

Mitigation: Mars-radiation-rated compute where possible; ECC memory + TMR critical path; periodic restart cycles; in-habitat compute for non-time-critical operations.

Training data + distribution shift[2]

Impact: VLA models trained on Earth tele-operation data may not transfer to Mars conditions (gravity, dust, regolith). Edge cases multiply.

Mitigation: Mars-simulant training data; sim-to-real transfer techniques; on-Mars supervised fine-tuning; conservative confidence thresholds.

Crew workload for Mars supervision[3]

Impact: Mars-supervisor crew time is the second-most-expensive Mars resource (after life-support consumables). Supervision workload must be small fraction of crew time.

Mitigation: High auto-routing of routine decisions; only edge-cases reach human; supervisor shifts shared across crew; AI-assisted decision summarization for human review.

Alternatives & substitutes

Direct teleoperation (real-time Earth control)[3]

  • Maximum precision + judgment from Earth-side operators
  • No autonomy software development required
  • Familiar paradigm from Earth-based robotics
  • Impossible at Earth-Mars distance (8-48 min latency)
  • Real-time only at lunar distance (1.3 s round-trip)
  • Not viable architecture for Mars

When preferred: Lunar surface operations; never Mars.

Manual crew control (in-base human operators)[3]

  • Crew judgment in real-time
  • No autonomy software complexity
  • Direct interface paradigm
  • Crew labor is most expensive resource on Mars
  • Limits scale (one operator per robot)
  • Defeats the purpose of robots augmenting crew

When preferred: Critical-precision tasks; emergency override; never routine.

Requires

References

  1. Iverson, K., Maimone, M., Verma, V., Castano, R., et al. (2024). Mars 2020 Perseverance Rover: Autonomous Surface Mobility (ENav + AutoNav). NASA Jet Propulsion Laboratory, AIAA SciTech 2024. — Perseverance autonomous navigation (AutoNav + ENav) flight performance + algorithm description. 100 m/sol average with onboard hazard avoidance.
  2. Tesla Robotics + Figure AI + Apptronik + Agility Robotics (2024). Humanoid Robotics 2024: Optimus Gen 2 / Figure 02 / Apollo / Digit — Public Specifications and Industrial Deployments. Tesla / Figure / Apptronik / Agility public statements. — Tesla Optimus Gen 2 (Dec 2023 reveal), Figure 02 (BMW Spartanburg deployment Aug 2024), Apptronik Apollo (Mercedes-Benz pilot 2024), Agility Digit (Amazon warehouses 2024). Cross-referenced via public IAC + earnings call statements + industrial pilot data.
  3. Larson, W. J., & Pranke, L. K. (Eds.) (1999). Human Spaceflight: Mission Analysis and Design. McGraw-Hill. ISBN 978-0-07-236811-4. — Standard reference for crewed-mission engineering: EVA architectures, life support, mission design, system trades.
  4. Drake, B. G. (Ed.) (2009). Human Exploration of Mars: Design Reference Architecture 5.0. NASA Johnson Space Center, NASA SP-2009-566. NASA/SP-2009-566. — NASA Mars Design Reference Architecture 5.0; mission architecture, MAV reference designs, ISRU mass budgets.