Robots can master individual skills and still fail long workflows. This guide explains compounding error, partial observability, progress tracking, verification and recovery architectures for multi-step tasks.
Introduction
A robot that succeeds 95% of the time on one step sounds excellent. If a twenty-step task required twenty independent 95% successes with no recovery, the probability of completing the whole chain would be only about 36%. Real errors are not independent, which can make the situation even worse because one bad state contaminates everything that follows.
This is the long-horizon problem. It is less about discovering another grasp primitive and more about preserving task state, detecting incomplete steps, recovering from physical errors and replanning before a small deviation becomes irreversible.
Research date: August 12, 2026. Information verified from official sources available as of August 12, 2026.
Direct answer
A robot that succeeds 95% of the time on one step sounds excellent. If a twenty-step task required twenty independent 95% successes with no recovery, the probability of completing the whole chain would be only about 36%. Real errors are not independent, which can make the situation even worse because one bad state contaminates everything that follows.
Key findings
- Long-horizon reliability is multiplicative: strong individual skills can still produce weak end-to-end task completion.
- Partial observability and hidden state make a short-window reactive policy lose information needed several steps later.
- Progress-aware systems explicitly track completed and remaining subtasks rather than assuming the planned sequence happened correctly.
- Goal2Skill separates a high-level memory and recovery planner from a low-level VLA executor and reports 32.4% average success versus 9.8% for its strongest baseline on selected RMBench tasks.
- Recovery should be measured as a first-class capability, including how often the robot notices failure, whether it can return to a recoverable state and when it escalates to a human.
What to measure before treating the claim as deployment evidence
A reusable evidence checklist applied throughout this article.
| Signal | Useful evidence | Common mistake |
|---|---|---|
| Capability | Repeated task success with defined trials and resets | Judging one edited demonstration |
| Autonomy | Human intervention and teleoperation disclosed | Calling scripted or supervised behavior autonomous |
| Generalization | Unseen variable and adaptation budget stated | Using zero-shot without defining what was unseen |
| Reliability | Long runs, recovery and failure logs | Reporting only peak performance |
| Deployment | Customer workflow, uptime and support burden | Equating hardware shipment with productive use |
Not every row applies equally to research papers and public-market analysis; the article specifies the relevant evidence.
Why success compounds
The simple multiplication example is useful because it shows why demos hide the wrong denominator. If each of 20 steps succeeded independently with probability 0.95, total success would be 0.95 raised to the twentieth power, about 35.8%. Raising each step to 99% would still yield only about 81.8%.
Real workflows also include dependencies. A slightly misplaced object at step four can make steps seven and eight impossible. The system therefore needs to manage state quality, not just local action accuracy.
Planning is not enough because the world edits the plan
A task planner can generate the right symbolic sequence and still lose when the drawer sticks, the object slips or a human moves something. Long-horizon autonomy is a closed-loop problem: plan, execute, observe, verify, update.
Open-loop execution assumes the predicted world survives contact. Physical environments punish that assumption quickly.
Progress tracking turns history into state
A progress representation records which subgoal is active, which prerequisites hold and what evidence confirmed completion. It can be textual, symbolic, visual or embedded in a learned state. The key is that the next decision depends on actual execution rather than the original plan index.
Gemini Robotics ER 2 makes this explicit through continuous-video progress classification and moment finding. Trace-conditioned VLA research similarly uses receding-horizon replanning to predict the remaining plan from the current observation.
Recovery requires a map of reversible states
Some failures are easy to repair: regrasp a dropped object. Others are irreversible: pour the wrong ingredient, break a seal or knock a fragile object off the table. A recovery-aware planner needs to know which states can be restored and which require escalation.
The safest strategy is often to stop earlier. Uncertainty thresholds and precondition checks can prevent the robot from executing a risky next step when the current state is ambiguous.
Goal2Skill separates semantics from motor control
Goal2Skill uses a high-level VLM planner for structured task memory, decomposition, outcome verification and error-driven correction, with a low-level VLA for visuomotor execution. The authors report 32.4% average success on representative RMBench tasks versus 9.8% for the strongest comparison baseline.
The absolute success rate is a reminder of the difficulty. Improving from roughly one in ten to roughly one in three is substantial research progress and still far from the reliability expected in unattended production.
The metric should include intervention and recovery cost
Task success alone hides whether a human reset the scene after every mistake. Deployment metrics should include intervention frequency, mean time to recovery, fraction of failures detected, fraction recovered autonomously and the number of irreversible errors.
A slower robot that recovers may outperform a faster robot that needs a technician. Long-horizon economics depend on total completed workflows per human support hour, not peak action speed.
Limitations and missing information
- The step-independence calculation is an illustration, not a measured success model for any specific robot.
- Research benchmarks differ in task length, reset policy and failure definition.
- High-level recovery cannot correct mechanical limits or unsafe low-level control.
- Progress trackers can themselves be wrong, so critical state transitions need independent verification.
Conclusion
Long-horizon autonomy is where robotics systems stop being collections of impressive skills and start becoming workers. The deciding capabilities are memory, verification, progress tracking and recovery.
The most informative benchmark is no longer 'can the robot do this action?' It is 'how many complete workflows can it finish before a human has to intervene, and what does it do when the world refuses to match the plan?'
Frequently asked questions
Why do robot success rates fall on long tasks?
Every step creates another opportunity for perception, control or planning error. Errors can also compound because later steps depend on the physical state produced earlier.
What is progress tracking?
It is an explicit estimate of which subtasks are complete, what state the world is in and what remains to be done.
What is robot failure recovery?
The system detects that an expected state transition failed, chooses a safe corrective action or replans, and verifies the new state before continuing.
What metrics matter beyond task success?
Human interventions, detected-failure rate, autonomous recovery rate, time to recovery, irreversible-error count and completed workflows per support hour are useful deployment metrics.
Sources and methodology
Research was checked on August 12, 2026. Current-company claims use official company or government material where available, while financing and listing details are cross-checked with Reuters.
Research-paper performance numbers are attributed to the authors and are not treated as independent validation. Benchmarks with different robots, tasks, resets or success definitions are not ranked as if they were directly comparable.
Related TechniaHQRobot guides
Official image recommendations
Use the exact robot and generation named below. Confirm reuse rights with the source owner before publication or social distribution.
Fact-check report
Verified:
Share this article
Share the current TechniaHQRobot article page.
Continue reading
Open the latest robotics reporting, Physical AI analysis and hardware notes.