SLA model tops RoboDojo-Real robot leaderboard, doubling π0.5 success rate
机器之心 · wechat · 2026-11-12
HKUST(GZ) and OLADimensions propose SLA, a Spatial-Language-Action model that extends spatial supervision from post-training into continual pre-training, using a frozen 3D perception model (OmniVGGT) as a spatial teacher aligned to the VLA's intermediate visual features.
Key results
- Tops the RoboDojo-Real leaderboard: 42.3 avg score and 33.9% full-task success vs π0.5's 22.9 and 12.8% (2.6x)
- 83.8% object-level success across eight real-world conditions, +24 points over π0.5
- Robust under moved targets, dynamic interference, and moving tabletops; adapts to a Franka arm with just 10 extra demos (78% success)
- RoboTwin Clean2Random: 72.3% success, leading all ablation combinations at every data scale (5%–100%)
Technical highlights
- Unifies six robot datasets (1.71M trajectories, 14,330 hours) via end-effector pose representation, unified coordinate conventions, and SO(3) relative rotation handling
- State dropout during training encourages visual reliance; spatial teacher removed at deployment with zero added inference cost
- Continual pre-training: 48 H100s for 110 hours
Paper, code, and project page are public.
More from Embodied
- GXO to deploy up to 1,000 Apptronik Apollo humanoid robots in warehouses over four years — CyberRobooo · 2026-11-13
- Waymo robotaxi spotted quietly obeying a 'no right turn on red' sign humans often miss — reed · 2026-11-13
- Walden Robotics unveils Sage (LBM 2.0): factory deployments reshape how robots scale — adnothing · 2026-11-13
- Blind Dexterity wins Best Paper at CoRL 2026 workshop: push-resilient humanoid walking using only joint encoders — _krishna_murthy · 2026-11-13
- Bambu Lab blocked a 34x traffic surge that nearly took its cloud down, meme follows — ngxson · 2026-11-13
- FLASH-ALOHA open-sources leader-arm-only teleop hardware, CAD, and code — philfung · 2026-11-13