← All Posts
EngineeringApril 2026·Shengjie Lin, Robert Sun

Introducing Instinct

Dexterity is the only company that has deployed Physical AI with the sense of touch and force control in production.

One tactile skill. Any task. Zero retraining.

Median Placement
4.55 seconds
Every placement guided entirely by feel
Sensing Rate
500–2,000 Hz
10–100× faster than human touch
Every Placement Unique
5–11
live decisions per placement — no two are the same
01

Touch Is the Missing Sense

Force sensing is not new. Robots have had force/torque sensors for decades. But reading a force is not the same as understanding it. The gap between sensing contact and acting intelligently on contact — adapting strategy, searching for openings, recovering from the unexpected, all at production speed — has never been closed in deployment.

Not a sensor. Not a compliance mode. An intelligence that operates through touch.

Instinct reads contact forces 500–2,000 times per second, reasons about what they mean, and acts. It is the first system to close the gap in deployment.

Superhuman reflexes. A human tactile nerve impulse takes 20–50 ms from fingertip to brain. Instinct reacts in 0.5–2 ms. 10–100× faster than human touch.
02

Proven in Production

No one has published production benchmarks for force-guided manipulation. Until now.

Instinct's flagship deployment is box packing into truck trailers. The robot must feel its way into constrained gaps between existing boxes, adjusting in real time based on what it touches. Every placement is guided entirely by feel.

The robot does not know the exact shape of the gap. It does not know how neighboring boxes will shift under pressure. It must discover these things through contact, react within milliseconds, and complete the placement reliably.

Here is what that looks like across a fleet of production robots.

Success Rate
81%
95% CI: 75%–86% · fleet-wide production
Median Time
4.55s
74% of placements under 6 seconds
Placements Analyzed
854
Random sample from our production robots
Placement Duration
Distribution of force-guided placement times across the production fleet
88% under 8 seconds. The vast majority of force-guided placements complete in under 8 seconds. The robot feels its way into a constrained gap, adapts its approach based on contact, and settles the box into position. At production speed.

When Instinct Cant Find a Way

The remaining 19% are not failures. They are decisions. Instinct detects that this particular gap cannot be completed safely and backs off — no damage, no collisions, no stuck states.

What happens next is what matters: the system recovers. Every single time. Instinct hands control back to Foresight, which retries the placement from a different angle, or selects an entirely different location. The box always gets placed.

This is a fundamental property of the architecture. Each capability is a composable transaction that can be retried, rerouted, or replaced without affecting the rest of the system. Instinct is not a standalone reflex — it is one layer in a stack that recovers by design.

03

What Touch Looks Like

Anatomy of a Placement
A single production placement — one robot, one box, one attempt
Calibrating
Approaching
Angling
Pressing
Settling

4.5-second placement — calibrate, approach, adjust angle, press, settle. Peak force: 51 N. A routine production placement.

In the clean placement, the robot calibrates, approaches, adjusts its angle, presses, and settles. 4.5 seconds, peak force 51 N. A routine production placement.

In the search placement, the robot presses but meets resistance. It sweeps laterally, feeling for an opening, adjusts angle, and presses through. 10 seconds, peak force 83 N. The Z axis shows the robot pushing upward to search for the slot — the primary packing direction is forward.

In the extreme placement, the robot presses for 6 seconds, reverses tilt, presses again, then searches laterally for nearly 4 seconds. Contact forces reach 123 N. After 14 seconds of persistent, adaptive effort, it finds a way.

The Robot Adapts

Instinct does not follow a script. In production, this is what adaptation looks like:

Change angle mid-placement
60%
Tries one approach angle, feels resistance, reverses to try another.
Search for an opening
15%
Sweeps laterally, feeling for a gap it can press through. Emergent search driven by touch.
Pause and recover
18%
Detects an unexpected obstacle, pauses, then resumes with a modified approach.
The simplest placement takes 5 decisions. The most complex takes 11. No two are the same.
Where Time Goes
How the robot spends its time during a force-guided placement
Pressing through contact — 47%. Moving into the gap, guided by force feedback.
Searching through contact — 25%. Sweeping, reversing, feeling for an opening under force.
Sensing & preparation — 16%. Calibrating and sensing before committing.
Settling into position — 10%. Final seating of the box.
04

Evaluations

Production data is the ultimate evaluation. What follows are benchmarks from a random sample of force-guided placements across a fleet of production robots, with distribution-free confidence intervals.

Force Precision
Peak contact force as a percentage of the safety limit — how much headroom does Instinct maintain?

At the median, Instinct uses 20% of the available force envelope. At the 90th percentile, 32%. Even at the 99th percentile — the hardest placements in the fleet — it stays below 60%. This is not a system that brute-forces contact. It is precise.

Difficulty Scaling
Success rate by placement complexity (number of decisions) — Wilson score 95% CIs
Failures are fast. When a gap is physically impossible, Instinct detects it quickly and backs off. Longer placements mean the robot is actively searching for the best spot — and when it searches, it almost always finds a way.
Operating Envelope
Every production placement: duration vs peak contact force. Green = success, red = graceful abort.

Each dot is one real production placement. The operating envelope is wide — Instinct handles placements from 1 to 22 seconds, with forces from near-zero to 244 N — and the successes and failures are cleanly separated. Failures cluster at low durations (fast detection of impossible gaps) and the system never exceeds its safety envelope.

Cross-Robot Consistency
Success rate per robot with Wilson score 95% confidence intervals

Instinct is not tuned to one robot. Across the production fleet, confidence intervals overlap — the skill transfers to different hardware with consistent performance. Same code, different arms, same results.

05

One Skill, Any Task

The deepest test of any capability is whether it transfers. A system that works on one task proves engineering. A system that works on any task — with different objects, constraints, and physics — proves a general capability.

Instinct transfers because it operates on contact physics, not task-specific knowledge.

ProductionBox Packing into Truck Trailers

Constrained gaps, guided entirely by touch. Hundreds of placements per shift.

ProductionBox Placement in Totes

Packing boxes into totes with less than 1 cm of clearance. The robot feels its way into an impossibly tight fit.

DemoChair Stacking

Picks up chairs and stacks them vertically, feeling contact to nest each one. Same code, different object.

DemoTwo-Arm Tire Pickup

Both arms feel the tire from opposite sides, grip cooperatively, and place it. True bilateral tactile coordination.

Contact physics doesnt care what the object is. A cardboard box, a plastic chair, a rubber tire — each has different friction, stiffness, and contact behavior. Instinct generalizes across all of them because it operates at the level of force and reaction, not object identity.
06

Instinct and Foresight

Instinct does not operate in isolation. It is deeply integrated with Foresight, Dexterity’s world model. The relationship is bidirectional.

Foresight decides where and when to deploy Instinct. The world model reasons about the scene, evaluates which gaps are feasible, and triggers the tactile system when force-guided manipulation is needed. In a typical production shift, Foresight triggers Instinct over two thousand times.

In the other direction, Instinct tells Foresight what the world actually feels like. As the robot presses a box into position, the forces carry information: the gap is narrower than expected, a surface is angled differently than it appeared on camera, a neighbor shifted. This contact data flows back to Foresight, which updates its understanding to match reality.

Think of it as walking around a room at night. Your eyes gave you a map before the lights went out, but your hands tell you where things actually are. You don’t just navigate by touch — your mental model of the room updates with every contact. That is what Foresight does with Instinct’s data.

Foresight
World Model
Decides where to act
Predicts outcomes
~2,000 triggers/shift
Plan & Context
Contact Reality
Instinct
Tactile Intelligence
Feels and reacts
Adapts in real time
Up to 2 kHz · <2ms

This is what separates Instinct from laboratory force control. A standalone controller is a reflex — useful, but limited. Instinct, integrated with a world model that understands context and predicts consequences, is intelligence.

07

Touch Goes Everywhere

Every task where a robot must feel its way through contact is now in reach. Assembly. Fragile goods. Irregular objects. Dense packing. Anywhere contact matters more than coordinates, Instinct applies.

Touch is no longer the missing sense.

Dexterity • April 2026