Skip to content

Robotics & Autonomous Manufacturing

  • Exit capability

    Build integrated perception-planning-control systems that manipulate the physical world and survive contact, uncertainty, and production constraints.

  • Frontier targets

    Humanoid robots; autonomous factories; eldercare robots; hazardous-work robots; orbital/lunar robots.

  • Choose this track if

    Best if you want intelligence embodied in machines.

Kinematics

Week 1: Robot anatomy and frames

Reading/source: MIT Robotic Manipulation Fall 2025: anatomy + Ch. 2 hardware + Ch. 3 frame notation

Know: Represent robots, frames, joints, end effectors, sensors and actuator interfaces precisely.

Reconstruct: Regenerate homogeneous transforms, composition/inverse and twist intuition.

Do: Model a 6-DOF arm in simulation; verify frame transforms against visual geometry.

Context: MIT's 2025 manipulation course explicitly integrates perception, planning and control in unstructured environments.

Defend: Which errors are coordinate mistakes versus physical-model mistakes?

Gate: Pass if every transform has declared source/target frame and unit tests.

Source: source

Week 2: Forward/inverse kinematics

Reading/source: Robotic Manipulation Ch. 3; Modern Robotics kinematics chapters

Know: Solve forward/inverse kinematics and recognize multiplicity, singularities and unreachable poses.

Reconstruct: Derive manipulator Jacobian from differential kinematics.

Do: Implement numerical IK with joint limits and multiple initial guesses.

Context: Modern Robotics provides a free formal treatment alongside MIT's applied stack.

Defend: What does a singularity mean operationally for the task?

Gate: Pass if solver detects/reports infeasible/singular cases rather than silently failing.

Source: source

Planning

Week 3: Trajectory generation and optimization

Reading/source: Manipulation Ch. 3.5-3.9 + Wave-1 optimization

Know: Generate smooth feasible trajectories with velocity/acceleration/joint constraints.

Reconstruct: Derive cubic/quintic boundary-condition trajectory coefficients or equivalent optimization form.

Do: Plan pick-and-place trajectories and compare interpolation vs optimization-based paths.

Context: Optimization becomes geometry in configuration space.

Defend: Why can a smooth end-effector path produce violent joint motion?

Gate: Pass if trajectory respects all declared kinematic limits.

Source: source

Perception

Week 4: Cameras, geometry and registration

Reading/source: Manipulation Ch. 4 geometric perception

Know: Use pinhole-camera geometry, depth, point clouds and rigid registration.

Reconstruct: Derive perspective projection and least-squares rigid alignment intuition.

Do: Register noisy point clouds; quantify pose error under occlusion/outliers.

Context: Perception is not a preprocessing box; pose uncertainty must propagate into planning.

Defend: What uncertainty does your perception pipeline actually output?

Gate: Pass if planning consumes uncertainty, not only a point estimate.

Source: source

Week 5: Learned perception

Reading/source: Manipulation perception chapters + modern segmentation/detection selected paper

Know: Combine geometric priors and learned features without confusing benchmark accuracy with task utility.

Reconstruct: Reconstruct cross-entropy/pose-loss objective and calibration concept.

Do: Train/fine-tune a small visual model on synthetic clutter; evaluate downstream grasp success.

Context: Modern robotics uses deep perception but still needs geometric and physical consistency.

Defend: Which perception errors matter to manipulation and which do not?

Gate: Pass if evaluation includes downstream task loss and out-of-distribution clutter.

Source: source

Dynamics

Week 6: Rigid-body dynamics

Reading/source: Underactuated Robotics: multibody dynamics chapters

Know: Derive manipulator equations and understand inertia, Coriolis, gravity and contact forces.

Reconstruct: Regenerate M(q)qdd+C(q,qd)qd+g(q)=Bu+Jᵀλ structure.

Do: Simulate torque-controlled 2-link/underactuated mechanism; compare model and numerical integration.

Context: Underactuated notes emphasize dynamics as central to agile robotics.

Defend: Which terms are coordinate-dependent and which reflect physical energy?

Gate: Pass if energy/momentum checks expose at least one implementation bug.

Source: source

Control

Week 7: Trajectory tracking and impedance

Reading/source: Underactuated + Wave-1 control; manipulation contact-control material

Know: Design feedback that tracks motion while remaining compliant under contact/model error.

Reconstruct: Derive computed-torque/PD linearization and impedance relation.

Do: Control an arm against uncertain object stiffness; compare position and impedance control.

Context: Contact-rich manipulation makes rigid trajectory tracking insufficient.

Defend: When should the controller deliberately allow error?

Gate: Pass if contact forces remain bounded under stiffness mismatch.

Source: source

Contact

Week 8: Friction, grasping and contact mechanics

Reading/source: Manipulation contact/grasping chapters

Know: Model unilateral contact, friction cones, grasp wrench space and force closure.

Reconstruct: Derive planar friction cone and simple force-closure criterion.

Do: Optimize contact forces for a grasp; perturb friction coefficient/object mass.

Context: Dexterous manipulation is often constraint/contact management.

Defend: Why can a kinematically valid grasp be dynamically impossible?

Gate: Pass if robustness is evaluated across friction/mass uncertainty.

Source: source

Planning

Week 9: Collision-free motion planning

Reading/source: Manipulation motion-planning chapters + LaValle/standard planning concepts

Know: Own configuration space, sampling-based search and optimization-based motion planning.

Reconstruct: Reconstruct RRT/PRM algorithm and probabilistic completeness intuition.

Do: Implement RRT or use Drake planner in clutter; compare path quality/runtime/failure.

Context: Planning difficulty grows with dimension and narrow passages.

Defend: What does planner failure tell you about feasibility?

Gate: Pass if algorithm reports uncertainty/timeout and uses deterministic replay seeds.

Source: source

Week 10: Task and motion planning

Reading/source: Manipulation task-and-motion chapters

Know: Integrate symbolic task choices with continuous geometric feasibility.

Reconstruct: Formalize discrete task graph with continuous feasibility oracle.

Do: Plan a multi-object rearrangement task where greedy ordering fails.

Context: Real household/industrial tasks mix logic and geometry.

Defend: Where should backtracking occur when geometry invalidates a symbolic plan?

Gate: Pass if system recovers from an intentionally impossible subplan.

Source: source

Uncertainty

Week 11: Planning under uncertainty

Reading/source: Manipulation uncertainty + Wave-1 probability/control

Know: Represent belief over state and choose information-gathering actions.

Reconstruct: Derive Bayes-filter predict/update and belief-space objective intuition.

Do: Implement partially observed pick/search task; compare certainty-equivalent vs belief-aware behavior.

Context: Uncertainty is often actionable: move camera, probe, regrasp.

Defend: When is sensing itself the best action?

Gate: Pass if active perception reduces task failure on held-out scenes.

Source: source

Replication

Week 12: Manipulation stack reproduction

Reading/source: MIT Fall 2025 assignments/schedule: simulation -> pick/place -> perception -> planning

Know: Integrate first-half course stack and reproduce a complete manipulation benchmark.

Reconstruct: Rebuild architecture/dataflow from blank page.

Do: Robot arm autonomously picks target among clutter in simulation; publish traces and failure taxonomy.

Context: MIT course assignments explicitly build a manipulation software stack.

Defend: Which subsystem dominates failure, and how do you know?

Gate: Replication Gate: reproduce ≥3 scene families with fixed evaluation seeds.

Source: source

Learning

Week 13: Imitation learning

Reading/source: CS285 imitation-learning material + robot trajectories

Know: Learn policies from demonstrations while diagnosing covariate shift.

Reconstruct: Derive behavioral-cloning likelihood/loss and DAgger distribution-shift argument.

Do: Collect scripted/expert demos in simulation; compare BC vs iterative data aggregation.

Context: Imitation is attractive for robotics but compounds errors off demonstration manifold.

Defend: What states does the learner visit that the demonstrator never labeled?

Gate: Pass if held-out recovery scenarios are included.

Source: source

Week 14: Reinforcement learning for robotics

Reading/source: CS285 + Underactuated learning/control discussions

Know: Use RL where models/controllers are inadequate without discarding safety/model knowledge.

Reconstruct: Reconstruct actor-critic objective and model-based/model-free tradeoff.

Do: Train policy for a contact task; compare sample efficiency to model-based baseline.

Context: Learning control remains difficult to reproduce and deploy robustly.

Defend: What capability did learning add that conventional control lacked?

Gate: Pass if baseline is strong and compute/sample cost is reported.

Source: source

Week 15: Sim-to-real and domain randomization

Reading/source: Robotics literature + manipulation simulation stack

Know: Treat simulator mismatch as a distribution problem, not magic transfer.

Reconstruct: Formalize parameter distribution and robust objective across environments.

Do: Randomize masses/friction/delay/sensors; evaluate on unseen fixed 'real' parameter set.

Context: Simulation scale is useful only if uncertainty covers relevant reality.

Defend: When does randomization hide ignorance rather than model it?

Gate: Pass if parameters are physically justified and worst-case failures inspected.

Source: source

Systems

Week 16: Real-time software and hardware architecture

Reading/source: Manipulation hardware station + embedded/control core

Know: Design timing, communication, state estimation and safe command interfaces.

Reconstruct: Create end-to-end latency/jitter budget and watchdog state machine.

Do: Build a simulated hardware abstraction with sensor dropout, delayed commands and emergency stop.

Context: Production robots fail in interfaces/timing as much as algorithms.

Defend: What happens if the perception process freezes while actuator loop continues?

Gate: Pass if all single-process failures transition to bounded state.

Source: source

Manufacturing

Week 17: Robot design for reliability and service

Reading/source: Wave-2 manufacturing/reliability + current robot architecture

Know: Move from prototype mechanism to field-maintainable product.

Reconstruct: Derive MTBF/MTTR availability and serial-system reliability.

Do: Choose one actuator/joint; produce BOM, tolerance, thermal, cable, bearing and service-life model.

Context: Humanoid/industrial economics depend on uptime and repair, not demo success.

Defend: Which component sets fleet availability?

Gate: Pass if maintenance labor/spares are quantified.

Source: source

Week 18: Autonomous production cells

Reading/source: Wave-2 manufacturing + robotics integration

Know: Design robots as elements of a production system with buffers, inspection and recovery.

Reconstruct: Regenerate Little's Law and bottleneck throughput relation.

Do: Simulate robotic cell with machine failures/rework; optimize throughput vs redundancy.

Context: Autonomous factories are operations systems, not collections of robots.

Defend: Where does autonomy create new queueing/quality failure modes?

Gate: Pass if line recovers from blocked station without human reset.

Source: source

Research

Week 19: Failure mining

Reading/source: All prior traces

Know: Build a mechanism-level taxonomy of manipulation failures.

Reconstruct: Construct causal graph from perception/contact/planning/control/hardware errors to outcome.

Do: Collect 100 failures in randomized scenes; label root mechanisms with inter-rater check.

Context: Research agendas should be driven by stable failure clusters.

Defend: Are your 'root causes' actionable or just subsystem labels?

Gate: Pass if taxonomy predicts at least one unseen failure.

Source: source

Week 20: Reproduce a current manipulation result

Reading/source: Choose an open recent manipulation paper supported by public code/data

Know: Learn paper reproduction and fair baseline comparison.

Reconstruct: Reconstruct claimed contribution and ablation logic.

Do: Reproduce central metric at reduced scale; document deviations.

Context: MIT course explicitly emphasizes reviewing active research papers and final projects.

Defend: Which conclusion survives when compute/data is normalized?

Gate: Extension Gate: one reproduction + one negative/adversarial condition.

Source: source

Week 21: Independent mechanism extension

Reading/source: Your reproduced paper + adjacent literature

Know: Propose smallest change that targets identified failure mechanism.

Reconstruct: Write falsifiable theory of change and expected failure conditions.

Do: Implement extension with preregistered held-out scene distributions.

Context: Novel robotics work should change behavior in physically meaningful conditions.

Defend: Why should this intervention generalize beyond your benchmark?

Gate: Pass if it improves targeted cluster without degrading unrelated tasks excessively.

Source: source

Systems

Week 22: Robot safety and human interaction

Reading/source: Wave-2 system safety + manipulation deployment context

Know: Integrate force/speed limits, safe states, uncertainty, human proximity and recovery.

Reconstruct: Derive kinetic-energy/safe-distance toy bound and hazard-control structure.

Do: Red-team robot with perception spoof, dropped object, human intrusion and stuck actuator.

Context: Home/hospital/factory robots operate around people and expensive assets.

Defend: What failure must be impossible versus merely unlikely?

Gate: Pass if safety constraint remains outside learned-policy authority.

Source: source

Capstone

Week 23: Integrated autonomous robot/factory system

Reading/source: All track sources

Know: Demonstrate perception -> planning -> control -> recovery -> operations loop.

Reconstruct: Regenerate architecture, timing, uncertainty and safety boundaries.

Do: Capstone: manipulation or mobile-manipulation system performing a multi-stage task under randomized failures.

Context: The capstone is judged on robustness/recovery, not best-case video.

Defend: Where does the system depend on human cleanup or hidden initialization?

Gate: Systems Gate: independent evaluator runs unseen scenes and failure injections.

Source: source

Week 24: Technical design review and scale plan

Reading/source: NASA/industry-style review mindset + manufacturing core

Know: Turn prototype into research/product roadmap with cost, reliability and learning plan.

Reconstruct: Produce requirements, interfaces, verification matrix and top 10 risks.

Do: Final report/video/reproducible code + 10k-unit manufacturing/maintenance sketch + 12-month research agenda.

Context: The frontier is useful robotics at acceptable cost and uptime.

Defend: Which bottleneck would you spend the next $1M and 10 engineer-years on?

Gate: Capstone Gate: mechanics/control reviewer + ML reviewer + manufacturing/safety reviewer.

Source: source

Research gates

G1 Replication

Required performance: Reproduce an open manipulation/planning/control result across multiple scene seeds.

Minimum artifacts: Simulation files; metrics; videos; failure taxonomy; baseline implementation.

Pass criterion: Best-case video is insufficient.

G2 Extension

Required performance: Improve one mechanism-level failure cluster without hiding regressions.

Minimum artifacts: Held-out scenes; ablation; physics/control explanation; compute/data normalization.

Pass criterion: Must survive randomized contact/perception/model conditions.

G3 System Closure

Required performance: Run perception->planning->control->recovery with timing/hardware abstractions and safety.

Minimum artifacts: Latency budget; safe states; fault injection; maintenance/reliability model.

Pass criterion: Independent evaluator chooses unseen scenes/failures.

G4 Research Defense

Required performance: Translate robot prototype to useful deployed fleet/production cell.

Minimum artifacts: Design review; BOM/cost; uptime/spares; verification; 10k-unit or fleet plan.

Pass criterion: Must identify the dominant uptime/economics bottleneck.

Frontier technologies primarily routed here