Dominic Rigby

KinetIQ Ascend: Toward 100% Reliable Manipulation and Superhuman Speed

Date: 28th October 2026

Blog Post

Related: RL Token: Bootstrapping Online RL with Vision-Language-Action Models

π0.5: a Vision-Language-Action Model with Open-World Generalization

A Visual Dive into Conditional Flow Matching

Key Points

• Humanoid goal: 99.9% success at superhuman speed. Demonstration data will not take us there… we need RL.
• End-to-end learned system: sensor inputs to motion
• Large performance improvements observed post RL
• Current bottleneck is manipulation:
  • Combines perception, planning, continuous control and difficult to model envs
  • Small errors = large change in outcome
  • Contact rich: material behave differently

High Level of Solution

• Trained on demos first -> right direction but fails to generalise. Can’t handle edge cases
• End-to-end RL increases both speed and success rate.
• RL runs both in simulation and in real life

Behavioural Cloning

• Limited by human speed and quality of data
• Tried speeding up data, but is limited by physical limits. E.g. gravity doesn’t speed up
• Doesn’t handle edge cases well

RL

• RL is run both in sim and in real life
• RL is on VLAs
• Observed that applying RL to only the hardest part of the problem improves performance on whole task
• Observed that the policy maintained ability to do tasks it hadn’t practiced -> transferable skills
• Observed skills transferred; performance went up on other tasks after training on one.

RL Algorithm

• Conditional Flow Matching: strong and sample efficient at complex multi-modal distributions
• PPO based Flow-SDE approach:
  • Other approaches either too unstable or sample inefficient (e.g. criticless GRPO)
  • Overcoming prob density issues:
    • PPO paired with SDE comparison
    • Inject noise into all steps apart from last to add exploration
  • Regression based critic to help with credit assignment
    • Authors noted it was surprisingly easy
    • Critic had independent weights but from same beahvioural cloning checkpoint -> make me assume they changed whole model

Architecture

• Images -> Vision encoder
• Text, robot state and actions -> tokeniser
• Denoising process cross-attends to tokens

Action Generation

• Action chunk far longer than update period: 2 second action chunk, 0.2 second update period
  • Allows us to absorb latency
  • If latency increases, we just continue to execute actions
• Action prefixes:
  • To encourage continuity, they feed the actions which will execute during generation into the model
  • There is a loss term to replicate these actions at the start of our sequence
• Asynchronous inference:
  • Actions generate whilst last action is executing
  • Action pre-fix is the actions which will execute whilst we generate

Reward Design

• Dense rewards require too much tuning
• Sparse binary reward: for success is utilised
  • No term for time, only discount encourages speed
• Penalise known bad behaviours such as misgrasphs
• R = r_task_sucess + \alpha * r_recoverable_error

Non-Stationary Real Environments

• Real environment changes over time
  • Operators change who understand rules differently
  • Time of day changes: brightness etc
  • Robot wear and tear
• Comparing to stationary baseline makes robot performance look better
• 10% of rollouts are with baseline policy on current env

Speed Curiculum

• Gradually increase command frequnecy throughout training

Issues to Fix:

1. Sampler-Trainer Gap:
   • Robots do inference onboard with Nvidia Jetson Thor
   • Difference in kernels and numerics mean probabilities are different on the cloud -> introduces instabilities
   • Partial fix is to recalculate probabilities on the cloud
2. Jitter: fixed with action pre-fixes
3. Inference speed: fixed with asynchronous inference