Embodied intelligence / off-world systems
Physical AI for machines that leave Earth.
We study how robots perceive, move and act when the ground, gravity and signal can’t be taken for granted.
Explore the research Better models. Fewer assumptions.
Perception → Action → Adaptation
01 / Research directions
A foot slips. Dust covers a lens. A command arrives late. We work on the gap between a policy that looks good in simulation and a machine that can recover in the field.
Embodied foundation models for manipulation. We combine vision, joint state and contact history to study tool use with unfamiliar objects.
Open question Can a shared representation distinguish a secure grasp from a convincing image of one?
Legged and humanoid control across 0.16–1.0 g in simulation. Lower weight does not mean lower inertia. We study contact timing, balance and recovery after a missed step.
Autonomous rover navigation through loose soil, hard shadows and repetitive terrain. Slip estimation must update the route before the wheels dig in.
Sim-to-real transfer and radiation-tolerant edge inference. We study quantized policies, fault injection and bounded fallback behavior under a 15 W compute budget.
Human–robot teaming for habitat inspection and maintenance. Explicit uncertainty, legible intent and useful handovers when a task exceeds the robot’s authority.
02 / Concept platforms
Three platforms to make the research concrete. These are engineering concepts; all figures below are design targets, not validated or flight-qualified performance.
K–01 / Surface mobility
A patient terrain scout. Six driven wheels and passive rocker-bogie suspension keep the perception problem close to the ground.
DESIGN TENSION More traction costs energy. More speed costs certainty.
L–02 / Habitat operations
A humanoid for spaces built around people. Reach, bracing and two-handed tool use come before speed.
DESIGN TENSION A stable stance here may be an unstable impulse elsewhere.
O–03 / In-habitat inspection
A free-flyer for enclosed microgravity environments. Inspect a surface, hold position, and leave the air as quiet as possible.
DESIGN TENSION Every correction spends propellant. Drift is sometimes the better move.
03 / How we work
A simulator is a hypothesis about the world. Field work tells us where it is wrong. We preserve those disagreements as data, then run the loop again.
Report the intervention rate.
The successful run rarely tells the whole story.
04 / From the notebook
Short lab notes. Unfinished ideas, engineering trade-offs and questions worth keeping. These are working notes, not academic papers.
A lower normal force changes the traction problem. Why a fixed slip threshold is a brittle shortcut at 0.16 g.
Start with a wheel on a slope. Reducing gravity reduces the normal load, but the soil response does not simply scale with it. Cohesion, grain size and sinkage matter. A friction coefficient alone hides too much.
Our next model should estimate slip from visual motion and wheel odometry together, then carry that uncertainty into the route cost. The unresolved part: separating actual slip from visual odometry drift in a nearly textureless scene.
STATUS / MODELING QUESTION · NO FIELD CLAIM
Recovery begins before the next foothold. A note on contact uncertainty, momentum and knowing when to brace.
A controller that waits for a clean contact label has already spent part of its recovery window. We are interested in policies that maintain several contact hypotheses and choose a useful action before certainty arrives.
The proposed comparison is deliberately small: a delayed contact signal, a missed foothold and a nearby handrail. Measure peak angular momentum, time to a stable stance and human interventions. A fast recovery that relies on an impossible grip is still a failure.
STATUS / EXPERIMENT DESIGN · SIMULATION FIRST
Inference is only part of the power budget. The case for measuring a complete perception-to-action cycle.
A compact policy can look inexpensive in isolation. Add sensor decoding, state estimation, memory traffic and a safety supervisor, and the useful number changes. We want to report energy per complete decision, alongside tail latency.
Fault tolerance creates a second trade-off. Redundant computation costs power; detecting an impossible action and entering a bounded hold state may cost less. The open question is how much fault coverage that simpler path can provide.
STATUS / MEASUREMENT PLAN · 15 W DESIGN BUDGET
05 / Open positions & collaboration
We’re a small independent lab. We welcome conversations with people working on robot learning, contact mechanics, embedded systems and field testing.
No formal vacancies are listed yet. For future roles or a concrete research collaboration, tell us what you’re trying to understand.
contact@actuceptlab.fyi A plain email. A human conversation.