Embodied AI & Deliberative Robotics

Memory, Cause-and-Effect Models, and Testing Actions Before a Robot Moves

Executive Overview

The autonomous robotics industry is advancing quickly in motor control, learned action skills, and task planning. However, commercial fleets operating in complex environments still face a fundamental limitation: they often react without first considering the wider situation and likely consequences.

The 3x3 Institute is developing a deliberative runtime that gives autonomous robots and fleets a lasting model of their surroundings, memory over time, the ability to test likely consequences before acting, and a record of why each action was taken.

Concretely, this is the difference between a robot that reacts well and a robot that thinks before it moves. A learned skill can handle the next ten milliseconds: it sees an obstacle and avoids it. What it cannot necessarily do is hold the thought “a person walked behind that pallet eleven seconds ago and I have not seen them come out.” It may have no memory of what is no longer visible, no model of what its own action would cause, and no record afterward of why it did what it did. The deliberative layer fills those three gaps: memory, consequence, and accountability.

Four Parts of an Autonomous Robot’s Decision System

DELIBERATION — Think Before Acting

min – lifecycle

Cause-and-Effect Model • Memory Over Time • Test Actions Before Moving • Decision Record

PLANNING — Organize the Task

sec – min

Break Work into Steps • Plan Recovery • Coordinate the Fleet

SKILLS — Respond to Immediate Conditions

~10 ms

Connect Perception to Action • Learned Skills • Reusable Movements

MOTOR CONTROL — Balance and Movement

ms

Joint Balance • Bipedal Locomotion • High-Frequency Force Control

What the Deliberative Layer Adds

  1. A Lasting Model of the Surroundings: Maintains a checkable, real-time record of physical objects, their locations, and environmental constraints.
  2. Testing Actions Before Moving: Runs “what-if” tests in software before sending physical commands to the robot’s lower control layers.
  3. Memory and Confidence Tracking: Tracks how old information is, whether sensors may have drifted, and how much weight to give each source.
  4. A Record for Safety Review: Uses tamper-evident event records to support traceability, record-keeping, and safety reviews, including requirements relevant to ISO 10218 and EU AI Act Article 12 where applicable.

Where It Can Be Applied

  • Logistics & Warehousing: Multi-vendor autonomous fleet coordination, preventing deadlock cascades and physical interference.
  • Autonomous Industrial Assembly: Closed-loop precision assembly interlocks with real-time anomaly isolation.
  • Defense & Unmanned Systems: Onboard deliberative reasoning for autonomous platforms operating in jammed or communication-degraded environments.

To discuss an application to your own systems, Get in Touch →  Explore the Methodology →