Robotics: January 2026 Week 1

1. Closing the Reality Gap: Zero-Shot Sim-to-Real Deployment for Dexterous Force-Based Grasping and Manipulation

arXiv:2601.02778

Why Novel: First demonstration of controllable force-based grasping on a multi-finger dexterous hand trained entirely in simulation and transferred zero-shot to real hardware. Eliminates need for torque sensors via current-to-torque calibration.

Key Innovations:

• Computationally efficient tactile simulation computing distances between dense virtual tactile units and objects via parallel forward kinematics
• Per-joint current-to-torque calibration that maps motor current to joint torque, eliminating torque sensor requirements
• Actuator dynamics modeling with randomization of non-ideal effects (backlash, torque-speed saturation) for robust transfer
• Asymmetric actor-critic PPO with full-state tactile-torque policy achieving force-controllable grasping and in-hand rotation

Evidence:

• fig:1 — Framework overview showing tactile + torque integration for dexterous manipulation
• fig:2 — Calibration and alignment of current-force (real) vs torque-force (sim) properties
• fig:6 — Visualization of force-adaptive grasping in real-world and simulation
• fig:8 — Grasping with controllable force magnitudes from low to high strength
• tab:4 — Ablation on observation combinations validating tactile+torque necessity

Impact: Provides practical, reproducible recipe for training full-state tactile-torque policies entirely in simulation for dexterous manipulation, enabling force-sensitive skills previously requiring extensive real-world training.

2. VLM4VLA: Revisiting Vision-Language-Models in Vision-Language-Action Models

arXiv:2601.03309

Why Novel: First systematic study revealing that VLM general capabilities are poor predictors of downstream VLA performance, and identifying the vision encoder (not language) as the primary performance bottleneck.

Key Innovations:

• VLM4VLA: minimal adaptation pipeline converting general VLMs to VLA policies with tiny parameter overhead for fair comparison
• Discovery that improving VLM performance on embodied auxiliary tasks (QA, pointing, depth) does NOT guarantee better downstream control
• Modality ablations proving visual module is the bottleneck, not language components
• Injecting control-relevant supervision into vision encoder yields consistent gains even when frozen during downstream fine-tuning

Evidence:

• tab:1 — Results on Calvin ABC-D showing VLM4VLA competitive with expert VLAs
• tab:2 — Results on SimplerEnv-Bridge and Libero-10 benchmarks
• fig:3 — Linear relationship analysis between VLM capabilities and VLA performance
• fig:4 — Box plots showing auxiliary task performance doesn't transfer to VLA control
• tab:4 — Impact of injecting action information into VLM vision encoder

Impact: Fundamentally changes how the field approaches VLA architecture design by showing VLM pretraining is necessary but insufficient, and directing attention to aligning visual representations with low-level control needs.

3. RoboReward: General-Purpose Vision-Language Reward Models for Robotics

arXiv:2601.00675

Why Novel: First comprehensive benchmark and training dataset for VLM-based robot reward models, demonstrating that accurate offline reward prediction translates to better online RL performance and that smaller specialized models can outperform much larger VLMs.

Key Innovations:

• RoboRewardBench: 2,800 human-verified real-robot episodes for standardized evaluation across diverse tasks/embodiments
• Negative examples data augmentation via counterfactual relabeling and temporal clipping to generate calibrated failures from success videos
• RoboReward 4B/8B models trained on 45K episodes outperforming frontier VLMs on short-horizon reward prediction
• Demonstrated strong correlation between offline reward accuracy and downstream RL performance

Evidence:

• fig:1 — Overview showing dataset construction and model training pipeline
• fig:3 — Strong positive correlation between reward accuracy and downstream RL performance
• fig:4 — Counterfactual relabeling approach for generating partial success/failure pairs
• tab:1 — RoboRewardBench results showing trained models outperform larger VLMs
• tab:2 — Real robot RL showing RoboReward 8B substantially outperforms Gemini Robotics-ER 1.5

Impact: Provides practical path to deploying VLM-based rewards for real-world robot RL, releasing data, models, and benchmark to advance general-purpose reward models for robotics.

1. InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation

arXiv:2601.02456

Mixture-of-Transformers architecture unifying scene understanding, visual foresight, and action execution achieves 14.5% improvement on daily tasks and 40-73% boost in dynamic settings over pi0 and GR00T.

2. CycleVLA: Proactive Self-Correcting Vision-Language-Action Models via Subtask Backtracking and Minimum Bayes Risk Decoding

arXiv:2601.02295

Enables VLAs to anticipate failures before they occur via progress-aware subtask monitoring and MBR-based test-time scaling for retries.

3. Value Vision-Language-Action Planning & Search

arXiv:2601.00969

Augments frozen VLA with lightweight value head enabling MCTS planning, achieving 5-6% improvement on LIBERO with reduced search simulations.

4. Contractive Diffusion Policies: Robust Action Diffusion via Contractive Score-Based Sampling

arXiv:2601.01003

Applies contraction theory to diffusion sampling, reducing solver/score-matching errors and improving data efficiency with minimal overhead.

5. Action-Sketcher: From Reasoning to Action via Visual Sketches for Long-Horizon Robotic Manipulation

arXiv:2601.01618

Introduces Visual Sketch as explicit spatial intermediate grounding language in scene geometry, enabling interpretable long-horizon manipulation with human-in-the-loop correction.

6. DemoBot: Efficient Learning of Bimanual Manipulation with Dexterous Hands From Third-Person Human Videos

arXiv:2601.01651

Learns dual-arm dexterous manipulation from single unannotated RGB-D video via MANO-based retargeting and residual RL with temporal segmentation.

7. SOP: A Scalable Online Post-Training System for Vision-Language-Action Models

arXiv:2601.03044

Fleet-scale actor-learner framework enabling online multi-task post-training of VLAs with near-linear scaling and hours-to-improvement cycles.

8. Learning Diffusion Policy from Primitive Skills for Robot Manipulation

arXiv:2601.01948

Skill-conditioned diffusion policy abstracting 8 primitive skills achieves SOTA on CALVIN and LIBERO via VLM-based skill routing.

9. A High-Fidelity Digital Twin for Robotic Manipulation Based on 3D Gaussian Splatting

arXiv:2601.03200

End-to-end pipeline from sparse RGB to planning-ready collision meshes via 3DGS, achieving ~90% real-world manipulation success with 0.83cm placement error.