SIMA 2 was trained on 9 commercial 3D games and 12 research environments, totaling over 400,000 hours of gameplay data. Unlike game-specific AI (like OpenAI's VPT for Minecraft), SIMA 2 learns transferable skills that work across different game engines, art styles, and mechanics.
More impressively, SIMA 2 can play games it's never seen during training through zero-shot transfer—applying learned skills to completely new environments. This section documents:
- Training games (high proficiency, 65%+ success rate)
- Tested unseen games (moderate proficiency, 25-50% success rate)
- Potential games (likely compatible based on mechanics)
Commercial Training Games (Official Partners)
These are the 9 commercial titles used for SIMA 2's primary training, with partnerships between DeepMind and game studios:
1. No Man's Sky (Hello Games)
Genre: Space exploration, survival, procedural generation
Release Year: 2016
Training Hours: 60,000+
What SIMA 2 Learned:
- Planetary navigation (flying, walking, jetpack)
- Resource gathering (mining minerals, harvesting plants)
- Base building (constructing shelters, power systems)
- Space flight (takeoff, landing, interstellar travel)
- Inventory management (organizing resources)
SIMA 2's Capabilities in No Man's Sky:
- Explore and catalog new planets (70% success rate)
- Gather resources for survival (85% success rate)
- Build basic bases (60% success rate)
- Engage in space combat (35% success rate—struggles with dogfighting)
Key Transfer Skills: 3D navigation in open worlds, procedural content recognition (every planet is different), multi-stage crafting systems
Demo Task: "Fly to nearest planet, gather 50 carbon, build a base"—achieved in 18 minutes with 72% success rate across 100 trials.
2. Valheim (Iron Gate Studio)
Genre: Survival, Norse mythology, co-op building
Release Year: 2021
Training Hours: 55,000+
What SIMA 2 Learned:
- Viking-style survival (hunting, farming, crafting)
- Combat against mythical creatures (trolls, draugr, bosses)
- Sailing and exploration
- Progressive technology tree (stone → bronze → iron → late-game metals)
- Weather and environment awareness (stamina in rain, cold damage)
SIMA 2's Capabilities in Valheim:
- Gather resources and craft tools (90% success rate)
- Build Viking longhouses (75% success rate)
- Hunt deer and boar (80% success rate)
- Defeat early bosses (Eikthyr: 55%, The Elder: 30%)
Interesting Behavior: SIMA 2 learned to repair tools before they break—an emergent behavior not explicitly taught. It monitors tool durability and proactively visits workbenches, showing planning ahead.
Demo Task: "Build a complete stone house with chimney"—achieved in 35 minutes with 68% success rate.
3. Goat Simulator 3 (Coffee Stain Studios)
Genre: Sandbox chaos, physics-based comedy
Release Year: 2022
Training Hours: 30,000+
What SIMA 2 Learned:
- Absurdist physics manipulation (ragdoll control, object interaction)
- Quest completion in non-linear environment
- Parkour and acrobatic movement
- Using ridiculous abilities (jetpack goat, tall goat, whale goat)
SIMA 2's Capabilities in Goat Simulator:
- Navigate the open world (95% success rate)
- Complete fetch quests (70% success rate)
- Cause property damage (100% success rate—it's a goat simulator!)
- Use advanced abilities effectively (45% success rate—struggles with timing)
- Complete timed challenges (40% success rate)
Why This Game Matters: Goat Simulator taught SIMA 2 to handle unpredictable physics. Many games have deterministic mechanics, but Goat Simulator's intentionally wonky physics forced SIMA 2 to develop robust recovery strategies when things go wrong.
Demo Task: "Find and headbutt 5 humans"—achieved in 8 minutes with 82% success rate.
4. Satisfactory (Coffee Stain Studios)
Genre: Factory building, automation, first-person
Release Year: 2019
Training Hours: 45,000+
What SIMA 2 Learned:
- Complex automation systems (conveyor belts, assemblers, power grids)
- Resource chain planning (raw materials → intermediate products → final goods)
- 3D factory layout design
- Efficiency optimization (reducing bottlenecks)
SIMA 2's Capabilities in Satisfactory:
- Set up basic mining operations (85% success rate)
- Build conveyor systems (75% success rate)
- Design efficient factories (50% success rate—better than random, worse than experienced players)
- Manage power grids (60% success rate)
Emergent Skill: SIMA 2 learned spatial planning—it arranges factory buildings to minimize conveyor belt length, without being explicitly taught this efficiency principle.
Demo Task: "Automate iron rod production"—achieved in 25 minutes with 71% success rate.
5. Teardown (Tuxedo Labs)
Genre: Voxel destruction, puzzle, heist planning
Release Year: 2022
Training Hours: 35,000+
What SIMA 2 Learned:
- Destructible environments (everything can break)
- Creative problem-solving (multiple paths to objectives)
- Time-pressure planning (set up route, then execute under time limit)
- Physics-based puzzles (using vehicles, explosives)
SIMA 2's Capabilities in Teardown:
- Navigate voxel environments (90% success rate)
- Use tools to break through walls (85% success rate)
- Plan optimal heist routes (45% success rate—struggles with multi-objective optimization)
- Complete challenges under time limits (40% success rate)
Why This Game Matters: Teardown taught SIMA 2 that environments are mutable. Most games have static worlds; Teardown requires thinking of walls as "temporary obstacles I can remove." This transfers to other games with destructible terrain.
Demo Task: "Break through three walls to reach the safe"—achieved in 12 minutes with 78% success rate (when no time limit).
6. Hydroneer (Foulball Hangover)
Genre: Mining, physics-based gold rushing
Release Year: 2020
Training Hours: 25,000+
What SIMA 2 Learned:
- Physics-based resource extraction (water pressure, gravity-fed systems)
- Equipment placement and optimization
- Economic management (buying tools, selling gold)
- Hydraulic system design
Demo Task: "Set up a functional gold mining operation"—achieved in 30 minutes with 64% success rate.
7. Wobbly Life (RubberBandGames)
Genre: Life simulation, party game, physics sandbox
Release Year: 2020
Training Hours: 20,000+
What SIMA 2 Learned:
- Social mini-games (racing, hide-and-seek, cooking)
- Vehicle control (cars, planes, boats)
- Job simulation (pizza delivery, postal service)
- Wobbly physics character control
Why This Game Matters: Wobbly Life taught multi-modal skill switching—rapidly changing between driving, walking, mini-games. Helps SIMA 2 adapt to games with varied gameplay segments.
Demo Task: "Deliver 5 pizzas to correct addresses"—achieved in 20 minutes with 71% success rate.
8. Space Engineers (Keen Software House)
Genre: Sandbox engineering, space construction, survival
Release Year: 2019
Training Hours: 40,000+
What SIMA 2 Learned:
- Complex engineering (building spacecraft from components)
- Zero-gravity navigation
- Power systems and resource management in space
- Creative problem-solving with limited resources
SIMA 2's Capabilities in Space Engineers:
- Navigate zero-G environments (85% success rate)
- Mine asteroids (80% success rate)
- Build functional spacecraft (45% success rate—often creates non-functional designs)
- Repair damaged systems (55% success rate)
Emergent Behavior: SIMA 2 learned to test designs incrementally—it builds a small prototype, tests if it works, then expands. Not explicitly trained, shows transfer of scientific method.
Demo Task: "Build a mining ship and collect 100 ore"—achieved in 45 minutes with 52% success rate.
9. Eco (Strange Loop Games)
Genre: Civilization building, ecology simulation
Release Year: 2018
Training Hours: 30,000+
What SIMA 2 Learned:
- Ecosystem management (don't over-harvest, replant trees)
- Collaborative building (designed for multiplayer)
- Technology progression (stone age → modern civilization)
- Long-term planning (prepare for meteor strike endgame)
SIMA 2's Capabilities in Eco:
- Gather resources sustainably (75% success rate at maintaining ecosystem)
- Contribute to team projects (when given specific subtasks)
- Advance through tech tree (60% success rate)
Why This Game Matters: Eco taught long-term consequences—if you clear-cut forests, you run out of wood later. SIMA 2 learned basic resource sustainability without being told explicitly.
Demo Task: "Gather 50 wood without destroying more than 10 trees"—achieved in 25 minutes with 69% success rate.
Research Environments (DeepMind Custom)
Beyond commercial games, SIMA 2 trained on 12 purpose-built research environments designed to teach specific skills:
- 3D Maze Navigation: Tests spatial reasoning and memory. Success Rate: 95%
- Object Manipulation Gym: Physics puzzles (stack blocks, open doors, push buttons). Success Rate: 88%
- Multi-Agent Coordination: Requires cooperating with 1-3 other AI agents. Success Rate: 65%
- Language-Grounded Tasks: Follow instructions like "Put the red cube on the blue platform." Success Rate: 92%
- Procedurally Generated Worlds: New environment every episode. Success Rate: 78%
- Long-Horizon Planning: Tasks requiring 20+ steps. Success Rate: 55%
- Specialized Skills: Additional environments for climbing, swimming, flying, stealth, combat, and crafting
Unseen Games (Zero-Shot Transfer Tests)
DeepMind tested SIMA 2 on games it had never seen during training to measure generalization:
ASKA (Survival Game)
Similarity to Training Data: Very similar to Valheim
Zero-Shot Success Rate: 42%
With 10 Minutes Demonstration: 65%
What Transferred Well: Resource gathering, building, crafting
What Failed: Game-specific mechanics (totems, ancestral worship)
Minecraft (via MineDojo Environment)
Similarity to Training Data: Moderate (voxel world like Teardown, but different mechanics)
Zero-Shot Success Rate: 25%
With 10 Minutes Demonstration: 40%
What Transferred Well: Mining, basic crafting
What Failed: Redstone (too complex), nether navigation (different environment)
Note: Despite no Minecraft training, SIMA 2 outperformed random actions. Shows genuine skill transfer.
Subnautica (Underwater Survival)
Similarity to Training Data: Low (underwater mechanics rare in training set)
Zero-Shot Success Rate: 18%
With 10 Minutes Demonstration: 35%
What Transferred Well: Resource gathering, base building
What Failed: Managing oxygen (new mechanic), underwater navigation (depth perception)
Cuphead (2D Platformer)
Similarity to Training Data: None (SIMA 2 trained only on 3D games)
Zero-Shot Success Rate: 3%
Why It Failed: Completely different perspective (2D vs 3D), requires frame-perfect timing, and pattern memorization (bosses have scripted attacks).
Lesson: SIMA 2 is a 3D generalist, not universal game AI.
Skills That Transfer Across Games
Core Transferable Skills
These work in almost every 3D game:
- WASD Navigation: Move forward/back, strafe left/right
- Mouse Camera Control: Look around, aim
- Jump: Navigate vertical obstacles
- Interact: Press E/F to open doors, pick up items
- Inventory Management: Understand grid-based inventory systems
- Health Awareness: Recognize low health, seek healing
- Objective Following: Find waypoint markers, track quest goals
Game-Specific Skills
Some skills are unique and don't transfer:
- Redstone logic (Minecraft)
- Spell combos (RPGs)
- Musical rhythm (rhythm games)
- Social deception (Among Us, Diplomacy)
Rule of thumb: Physical skills transfer better than conceptual/social skills.
How SIMA 2 Discovers Game Mechanics
When encountering a new game, SIMA 2 uses curiosity-driven exploration:
Phase 1: Basic Interaction (First 5 Minutes)
- Test all buttons: What does each key do?
- Observe UI: Health bar, stamina, inventory icons
- Try walking, jumping, interacting
Phase 2: Goal Inference (5-15 Minutes)
- Look for waypoint markers, quest text, highlighted objects
- Experiment: Try combining items, using tools on objects
- Form hypothesis: "This seems like a survival game. Probably need food/shelter."
Phase 3: Skill Application (15+ Minutes)
- Apply known strategies from similar games
- If Valheim-like: Try gathering wood, building shelter
- If shooter-like: Try finding weapons, cover
Phase 4: Refinement
- Learn game-specific mechanics through trial and error
- Update world model: "In this game, water restores health" (differs from Valheim where water drowns you)
Success rate correlates with similarity to training games:
- Very similar (ASKA ← Valheim): 65%
- Moderately similar (Minecraft ← voxel games): 40%
- Dissimilar (Subnautica ← mostly terrestrial games): 35%
- Completely different (2D platformer): 3%
Future Games: What Could SIMA 2 Handle?
Likely Compatible (Predicted 50%+ Success)
Games similar to training set:
- 7 Days to Die (Survival)
- ARK: Survival Evolved
- Raft (Ocean survival)
- The Forest
- Grounded (Honey I Shrunk the Kids survival)
Possibly Compatible (Predicted 25-50%)
Some shared mechanics:
- Terraria (2D but has 3D-like building)
- Don't Starve Together (survival, different perspective)
- Rust (PvP survival—social element is challenge)
- Deep Rock Galactic (co-op mining)
Likely Incompatible (Predicted Less Than 25%)
Very different mechanics:
- Civilization VI (turn-based strategy)
- FIFA 2024 (sports, requires tactics)
- Dark Souls (requires precise timing, pattern memorization)
- Visual novels (text-heavy)
- Rhythm games (musical timing)
Comprehensive Game Database
| Game | Genre | Training Hours | Success Rate | Key Skills Learned |
|---|---|---|---|---|
| No Man's Sky | Space exploration | 60,000 | 70% | Procedural navigation, space flight |
| Valheim | Viking survival | 55,000 | 75% | Crafting progression, boss combat |
| Goat Simulator 3 | Physics sandbox | 30,000 | 82% | Absurdist physics, quest completion |
| Satisfactory | Factory building | 45,000 | 71% | Automation, spatial planning |
| Teardown | Voxel destruction | 35,000 | 78% | Destructible environments, creative paths |
| Hydroneer | Gold rush mining | 25,000 | 64% | Physics-based extraction, hydraulics |
| Wobbly Life | Life sim | 20,000 | 71% | Multi-modal gameplay, social tasks |
| Space Engineers | Space engineering | 40,000 | 52% | Zero-G navigation, systems engineering |
| Eco | Ecology simulation | 30,000 | 69% | Sustainability, long-term planning |
| ASKA (unseen) | Survival | 0 (zero-shot) | 42% | Transfer from Valheim |
| Minecraft (unseen) | Voxel sandbox | 0 (zero-shot) | 25% | Transfer from voxel games |
| Subnautica (unseen) | Underwater survival | 0 (zero-shot) | 18% | Limited transfer |
What This Means for Game Developers
Opportunities
- Automated QA: SIMA 2 can playtest your game 24/7, finding bugs and soft-locks
- Difficulty Tuning: See how AI performs at different sections, adjust difficulty accordingly
- Content Discovery: AI may find unintended strategies or sequence breaks
- Accessibility: Train SIMA 2 on your game, offer it as an "AI assistant" mode for players who need help
Considerations
- Not Human-Level Yet: 65% success rate means it fails 1/3 of the time
- Requires 3D First-Person or Third-Person: Won't work for 2D or top-down games
- Training Data Needed: Zero-shot works but is weak. 5,000-10,000 hours of gameplay data for best results
- Social Mechanics Limited: Can't handle complex multiplayer diplomacy
Frequently Asked Questions
Can SIMA 2 play every game?
No. Currently limited to: 3D games (first-person or third-person perspective), real-time gameplay (not turn-based strategy), and primarily physical interaction (not text/dialogue-heavy).
What game should be added to SIMA 3's training?
Community suggestions DeepMind is considering: Minecraft (official partnership in progress), The Forest (high demand), Deep Rock Galactic (co-op mining), and Stardew Valley (different genre—farming sim).
Can SIMA 2 speedrun games?
Not competitively. It can complete games, but not optimize for speed like human speedrunners. Success rate matters more than speed.
Will SIMA 2 discover exploits/glitches?
Sometimes. In testing, SIMA 2 found: 3 previously unknown bugs in Valheim (reported to Iron Gate), unexpected building techniques in Satisfactory, and soft-locks in research environments. Not as good as dedicated bug hunters, but useful supplementary testing.
Can I train SIMA 2 on my own game?
Not yet—SIMA 2 isn't publicly available. But when/if released, expect requirements: 5,000+ hours of gameplay footage, API access or screen recording capability, and significant compute (100+ GPU hours for fine-tuning).
The Road to 100+ Games
DeepMind's internal roadmap (subject to change):
- 2025: 20-25 games (current state)
- 2026: 50 games (add more indie titles)
- 2027: 100 games (include competitive multiplayer)
- 2028: 250+ games (aim for "plays anything 3D")
Challenges remaining: Turn-based strategy (Civilization, XCOM), fighting games (frame-perfect timing), social deception (Among Us, Werewolf), creative games (The Sims, Little Big Planet), and narrative-heavy RPGs (Disco Elysium).
Each new genre requires architectural innovations. SIMA 2 is incredible at physical 3D action games. Expanding to other genres means new research breakthroughs.
Conclusion: A Versatile Game-Playing AI
SIMA 2's game repertoire demonstrates genuine generalization:
- 9 commercial games mastered (65-85% success rates)
- 12 research environments conquered
- 3+ unseen games handled with zero training (25-50% success rates)
What this proves: AI can learn transferable skills, not just memorize specific games. The same principles apply to robotics—an agent that masters diverse virtual environments can potentially master diverse physical tasks.
For gamers: SIMA 2 represents a future where AI companions can co-op with you in any game, provide tutorials by demonstrating techniques, and fill in for missing teammates.
For researchers: SIMA 2 is a testbed for embodied AI that's more accessible than robotics, more diverse than single-game AI, and more challenging than toy environments.
The next frontier: Games with social complexity (MMOs, MOBAs) and creative expression (Minecraft redstone, The Sims storytelling). These require moving beyond physical intelligence into social and creative AI.