Mysore sandalwood carving is a living knowledge system, not simply a collection of finished objects. Its value lies in how artisans read the grain, select a block, hold a tool, sequence cuts, correct mistakes and finish a surface without losing the wood’s fragrance or structural integrity. Any serious technology project must preserve that judgment and ensure artisans retain control over how their knowledge is recorded and used.
Reinforcement learning (RL) can help, but it should be applied carefully. RL systems learn by receiving rewards for actions in an environment. In this context, the environment could be a carving simulator, a robotic testbed or a digital material-planning tool. The objective is not to automate the artisan out of the process. It is to create safer practice environments, document tacit knowledge and support decisions that improve training, material efficiency and livelihoods.
Why Mysore sandalwood carving needs a practical preservation strategy
The craft faces connected cultural, economic and ecological pressures:
- Skills transmission is fragile: Apprentices need sustained access to master artisans, tools and suitable practice material.
- Raw material is regulated and limited: Sandalwood sourcing, storage, transport and sale involve legal and conservation requirements. A digital project must never encourage undocumented or illegal sourcing.
- Machine-made substitutes compete on price: Handmade work needs stronger provenance, storytelling and market access—not just more production.
- Knowledge is often undocumented: A finished carving does not reveal the sequence of decisions that produced it.
- Technology can extract value: Recording an artisan’s technique without consent, attribution or benefit-sharing can harm the very community the project claims to support.
A useful preservation programme therefore combines craft documentation, apprenticeship, responsible sourcing and commercial viability. AI is one layer in that system.
Where reinforcement learning can add real value
1. Build a risk-free practice environment
A training application could represent a virtual wood block with variables such as grain direction, density, moisture, tool angle and cut depth. An apprentice would choose an action—remove material, change tools, rotate the block or pause for inspection—and receive feedback based on surface quality, structural safety and material loss.
The reward function should be designed with artisans, not only engineers. A technically efficient cut may still be unacceptable if it damages a motif, ignores a traditional sequence or produces a finish that experienced carvers reject. The system should show *why* an action is recommended and allow the learner to compare its advice with a master’s explanation.
This approach can complement AI-powered games for learning programming: the same principles of progressive difficulty, immediate feedback and replayable practice can make craft education more engaging without turning it into a shallow game.
2. Capture tacit knowledge through demonstrations
The strongest dataset may come from annotated demonstrations rather than automatically generated actions. Record consenting artisans as they:
- inspect and grade a piece of wood;
- mark a design and establish reference points;
- select tools for different cuts;
- respond to grain changes or cracks;
- refine details and apply finishing techniques.
Video, audio, tool-motion data and written explanations should be linked to specific stages of a carving. Multiple artisans should be represented so the system does not mistake one person’s style for a universal rule. Local-language narration, including Kannada where appropriate, is essential for accurate knowledge transfer.
A searchable archive can preserve not only designs but also terminology, workshop practices, provenance and the artisan’s attribution. Like an AI tool for organising and preserving life stories, it should retain context and voice rather than reducing heritage to disconnected images.
3. Improve material planning and reduce waste
A reinforcement-learning or optimisation system could recommend how to orient a design within a legally acquired block of sandalwood. Inputs may include dimensions, grain observations, defects and the intended sculpture. Rewards can balance usable yield, structural strength, visual continuity and the artisan’s preferred working method.
This does not require a fully autonomous carving machine. A practical first product is a tablet tool that lets an artisan photograph or scan a block, mark defects and compare alternative layouts. The final decision remains human. Over time, anonymised workshop data can reveal recurring sources of waste and inform better teaching.
Material efficiency should be paired with compliance. The platform needs a chain-of-custody record, supplier documentation and clear rules for what data may be stored. It should also support alternatives for training, such as permitted substitute woods or synthetic blocks, so valuable sandalwood is not consumed during early practice.
Designing an ethical RL project in India
Start with a narrowly defined pilot
Do not begin by promising to digitise the entire tradition. Choose one measurable use case, such as training on basic tool control, documenting grain-selection decisions or planning small decorative panels. Define success in craft terms:
- apprentices make fewer unsafe or irreversible errors;
- master artisans judge the output as culturally and technically credible;
- material waste falls without lowering quality;
- learners complete more supervised practice;
- participating artisans receive payment, attribution and useful tools.
A small pilot with a Mysuru-based workshop, craft institution or artisan cooperative will produce better evidence than a large generic dataset.
Create a human-in-the-loop system
Master artisans should review demonstrations, label examples and approve learning content. Their feedback should be treated as a core signal in the reward model, not as a final marketing endorsement. If the AI recommendation conflicts with a craftsperson’s judgment, the interface should make that disagreement visible and record the correction.
The project team should include craft practitioners, conservation or legal experts, UX researchers, machine-learning engineers and people who understand India’s cultural and language context. An AI-guided software walkthrough can help non-technical workshop participants use the tool, but the interface must remain usable offline and on modest hardware.
Protect rights and access
Before recording or training a model, obtain informed consent in a language participants understand. Set out:
- who owns raw recordings and derived datasets;
- whether techniques can be licensed or sold;
- how artisans are paid for ongoing use;
- how attribution appears in the product;
- whether an artisan can withdraw material;
- which knowledge must remain private within a lineage or workshop.
Avoid publishing high-resolution designs or process data that enable cheap copying. Consider tiered access: public educational material, restricted professional content and private workshop knowledge. Any commercial revenue should include a transparent benefit-sharing mechanism.
A workable technical architecture
A modest first version could include:
1. Data layer: consented videos, photographs, tool metadata, multilingual transcripts and structured records of stages.
2. Simulation layer: a simplified 3D block, tool actions, collision rules and material-removal estimates.
3. Learning layer: imitation learning from artisan demonstrations, followed by RL in simulation. This is safer than letting an agent learn solely through physical trial and error.
4. Review layer: master-artisan scoring, explanations of recommendations and an audit trail of corrections.
5. Delivery layer: an offline-first mobile or desktop application that synchronises when connectivity is available.
Use synthetic or substitute materials for most simulation experiments. Physical robotics should come later, and only for constrained tasks such as tool-path testing under expert supervision. A high-throughput, low-cost inference setup may help deployment, but accuracy, explainability and data governance matter more than model size.
Measuring whether preservation is actually working
Track outcomes across four categories:
- Craft: quality ratings, motif fidelity, finish quality and preservation of regional styles.
- Learning: apprentice retention, supervised practice hours and improvement across repeated tasks.
- Resources: material yield, training waste, tool damage and energy use.
- Livelihoods: artisan income, attribution, new orders, cooperative participation and licensing revenue.
Do not use model accuracy as the headline metric. A system can predict tool movements well and still weaken the tradition if it narrows stylistic diversity or shifts control away from artisans.
A 12-month implementation roadmap
Months 1–3: form the artisan advisory group, document consent procedures, select one use case and audit sourcing and legal requirements.
Months 4–6: record demonstrations, create a multilingual taxonomy of tools and techniques, and build a simple searchable archive.
Months 7–9: develop the simulator, train an imitation-learning baseline and test reward criteria with apprentices and master carvers.
Months 10–12: run a supervised pilot, measure craft and livelihood outcomes, publish non-sensitive findings and decide whether RL adds enough value for expansion.
Conclusion
The best answer to how to preserve the craftsmanship of Mysore sandalwood carving through reinforcement learning is not to replace handwork with automation. It is to use RL selectively: to make practice safer, material planning more informed and tacit knowledge easier to pass between generations. The project succeeds only when artisans control representation, receive fair value and remain the final authority on craft quality.
For Indian builders, the opportunity is to create a responsible preservation infrastructure—local-language, offline-capable, consent-based and connected to real apprenticeships. Technology should strengthen the workshop, not flatten it into a dataset.