Three Capstones: From Engineering Skill to Independent Research

English · 简体中文

A capstone integrates knowledge, engineering, evaluation, and communication into reviewable work. It is not a demo video or a collection of connected repositories. Any “expert” conclusion remains bounded to the reviewed task and evidence.

C0 · Robot-loop diagnosis

• Prerequisites: M00–M04.
• Task: build or modify a small simulated loop connecting observation, state, control, and metrics.
• Required fault: inject a frame, timing, sensor, or actuator failure.
• Deliver: nominal and failed traces, logs, fault tree, safe stop, and reproduction receipt.
• Pass: 80/100 and at least one independent reviewer.

C0 demonstrates robot-loop engineering, not learned-policy quality.

C1 · Closed-loop learning system

• Prerequisites: goal-required modules from M00–M07, plus M10 and M11.
• Task: implement a simple baseline and a learning method under the same task, data, and evaluation contract.
• Required comparison: a no-learning/rule or state-only baseline and one learned policy.
• Required tests: held-out initial states, perturbations, failure stages, and at least one counterfactual ablation.
• Deliver: dataset card, training receipt, checkpoint, closed-loop result, negative results, and model card.
• Pass: 85/100 and at least one independent reviewer.

C1 demonstrates a controlled robot-learning system, not general embodied intelligence.

C2 · Independent reproduction and research extension

• Prerequisites: M00–M11 all passed.
• A — Independent reproduction: freeze the method, data, budget, and evaluation before reproducing a credible baseline; retain negative outcomes and differences.
• B — Original hypothesis: state a single-variable hypothesis that evidence can falsify.
• C — Controlled extension: compare baseline, change, and required ablations under matched budgets with multiple seeds and uncertainty.
• D — Distribution shift: test at least two axes among object/state, viewpoint, language, scene, latency, and embodiment.
• E — System loop: show how prediction/planning changes executed action and report safety, recovery, and stopping.
• Deliver: report, code, environment receipt, datasheet, raw metrics, checkpoints, video/traces, failure taxonomy, risk register, and defense material.
• Pass: 90/100 and at least two independent reviewers.

A successful reproduction alone cannot pass C2; a new method without a credible baseline cannot pass either.

Common capstone score (100 points)

Dimension	Points	Full-credit condition
Question and success contract	10	Task, I/O, metrics, boundary, and stop conditions are testable
Prerequisites and system design	10	Architecture choices and failure propagation are explained
Data and interfaces	10	Schema, synchronization, split, units, and embodiment identity are traceable
Baseline quality	10	Simple, budget-matched, and independently reproducible
Method implementation	10	Learner owns the key contribution and passes unit/interface checks
Closed-loop evaluation	10	Multi-episode task outcomes, stage metrics, and failures are complete
Ablation, OOD, and statistics	10	Counterfactuals, two shift axes, seeds, and uncertainty are adequate
Systems, safety, and recovery	10	Latency, limits, watchdog, recovery, and authorization boundaries are clear
Reproducible evidence	10	Commit, environment, data, config, commands, and raw artifacts are complete
Explanation and defense	10	Handles counterexamples and distinguishes fact, inference, hypothesis, and unknown

Any critical failure overrides the total score.

Example C2 questions

Direction	Falsifiable question	Required baseline and ablation
VLA	Does language change the correct action under the same visual state?	State/vision BC; correct, shuffled, and missing language
WAM	Does action-conditioned future prediction improve planning utility?	Model-free policy and WM+MPC; action swap, no action, no lookahead
Dexterity	Does a retargeting/contact representation improve task retention rather than only geometry?	IK/optimization baseline; geometry, contact, lift, and retention stages
Navigation/agents	Does recovery/replanning improve perturbed-task success without more collisions?	Planner without recovery; perturbation and recovery-component ablations
Sim-to-Real	Which identification or randomization factor reduces a declared transfer gap?	No-adaptation policy; single-factor randomization, HIL, shadow mode

Review and revision

Use the capstone review form. A reviewer must provide:

1. one counterexample that could overturn the conclusion;
2. one raw artifact personally checked;
3. the most serious unknown risk;
4. a pass, revision_required, or blocked decision.

Revision must preserve earlier results. A new experiment card records the reason and the single changed variable.

Hardware boundary

A capstone score never authorizes robot motion. Hardware requires site-, device-, operator-, and task-specific approval plus manufacturer limits and stop procedures. Learners without hardware may complete a simulation capstone, but must label the evidence level as simulation.