Meichen Liu’s Bipu turns messages into recorded robot-arm movements and electronic calls. Its useful Jev lesson is the boundary: interpret a message, let local policy restrict possible responses, then select from that menu—not generate arbitrary robot motion.

I inspected the Jev adapter, policy, runtime and controller backend and ran 44 offline tests successfully. I did not call live Jev, connect Telegram or move a robot. These tests cover software behavior and simulated controller states, not physical safety or interpretation accuracy; this is not a security audit.

Interpret first; restrict before choosing

  • Jev classifies messages into social intents such as praise, scolding, reassurance or a toy invitation, using recent interaction context.
  • Local code maintains energy, irritation and sleep, then limits the motion choices; the hardware backend further filters for reviewed transitions.
  • A second decision rates permitted motion/sound pairs. The program samples the validated distribution for expressive variety rather than always taking its top choice.
  • Low-confidence interpretation or an invalid response means waiting with an error—not silently switching to keyword rules.

Text cannot make the workspace ready

  • A message about toys cannot authorize the recorded block-play motion. The operator must explicitly confirm the props are positioned.
  • One play consumes that readiness; the operator must reset the arrangement before another.
  • The model cannot invent trajectories or write joint angles. The controller adapter checks recorded targets against configured limits and verifies preparation and final poses.
  • Stop cancels pending work but does not release motor torque. This is not an independent collision detector or emergency stop.

Read the demo’s evidence carefully

  • The author reports a real private Telegram message reaching Jev and a motion/sound choice in local preview; I did not reproduce that live path.
  • The pet film is edited from recorded performances; the octopus scene is scripted choreography, not proof of autonomous execution.
  • The author reports replaying individual motions on their arm, while autonomous physical transitions still require review.
  • The adapter’s interpretation threshold of 0.55 is an implementation setting, not demonstrated calibration. Typed output cannot establish a correct interpretation or safe movement.

Bipu is a concrete design example for bounded expressive agents. Its strongest evidence is inspectable policy and offline failure-path tests—not a validated autonomous robot.